manual_slop/conductor/tracks/fable_review_20260617/report.md

Fable System Prompt Review (Critical Analysis)

Track: fable_review_20260617 Initialized: 2026-06-17 Owner: Tier 1 Orchestrator (synthesis); Tier 2 Tech Lead (dispatch + QA) Status: IN PROGRESS (Phase 3)

What this is. A critical analysis of Anthropic's Claude Fable 5 system prompt (the public "Mythos" version), comparing it against Manual Slop's existing agent-directive corpus and Mike Acton's nagent patterns. Verdict framework: Useful / Persona Performance / Anti-User / Mixed.

Verdict summary. Fable is approximately [N]% useful, [M]% persona performance, [K]% anti-user. The full breakdown is in §0 (TL;DR Scorecard).

For the deferred nagent-rebuild. See decisions.md for the 15-20 concrete recommendations to apply when the user overhauls Manual Slop's agent directives (deferred 1-2 weeks per the user, 2026-06-17).

---

§0. TL;DR + Verdict Scorecard

Headline finding

Anthropic's Claude Fable 5 system prompt (the public "Mythos" version) is approximately 45% Genuinely Useful, 35% Persona Performance, 15% Anti-User Watch-Dogging, and 5% Mixed (by line-range weight across 10 cluster reviews). The useful patterns are concentrated in epistemic discipline, the memory/storage model, and the file-workflow rules. The persona patterns are concentrated in product branding, warm-tone framing, and mistake-handling theater. The anti-user patterns are concentrated in the user_wellbeing watch-dogging and the refusal-architecture safety theater. Manual Slop's per-developer, data-oriented, multi-provider, strict-HITL design is the philosophical opposite of Fable's consumer-product, persona-primary, vendor-locked stance; the two stances are mutually exclusive in their emphasis.

Top 3 adoptions (from decisions.md)

1. Adopt Fable's "search-default for current-state" rule (Fable System Prompt.md:158-164). Add to conductor/code_styleguides/rag_integration_discipline.md. The pattern is genuinely useful; Manual Slop's RAG is opt-in, but the "search before answering about current officeholders" rule should be default-on for current-state queries.
2. Adopt the "agents default to prose; use bullets only when asked or when content is multi-faceted" rule (Fable System Prompt.md:84-90). Maps directly to Manual Slop's AI-Optimized Compact Style at conductor/product-guidelines.md:39-49.
3. Adopt the "agents do not perform gratitude; they execute the task" rule (Fable System Prompt.md:124). Separable from its user_wellbeing context; useful as a tone rule for coding agents.

Top 3 explicit rejections (for AGENTS.md)

1. Reject Fable's "mental-health watchdog" framing (Fable System Prompt.md:92-124). Anti-user watch-dogging; the model is text generation, not a clinician.
2. Reject Fable's "Claude is deserving of respectful engagement" / "end_conversation tool when being mistreated" pattern (Fable System Prompt.md:154). Treats the model as a protected party with standing to demand dignity from the user.
3. Reject Fable's "Claude can share its concerns with the person openly" pattern (Fable System Prompt.md:108). The model has no concerns; it has a conversation log.

17-section verdict scorecard

§	Section	Source cluster	Verdict	One-line
§0	TL;DR + Verdict Scorecard	(summary)	n/a	This section.
§1	The 3 Sources	n/a	n/a	Fable vs Manual Slop vs nagent; what's in scope.
§2	The "Useful vs Persona vs Anti-User" Framework	n/a	n/a	The 4 verdict categories; why this lens.
§3	Fable's Product Branding & "Helpful Assistant" Persona	cluster 1	Persona Performance	Brand-specific noise; Manual Slop is per-developer, multi-provider.
§4	Fable's Refusal Architecture & "Safety Theater"	cluster 2	Anti-User + Persona (1 Useful caveat)	Refusal is a model attribute, not a directive; the audit-script layer makes refusals auditable.
§5	Fable's Mental-Health Watchdog Framing	cluster 3	Anti-User	The model is text generation, not a clinician; the strongest anti-user pattern.
§6	Fable's Tone & Formatting Constraints	cluster 4	Useful + Persona	Formatting discipline is useful; warm-tone persona is rejected.
§7	Fable's Mistake Handling	cluster 5	Persona + Anti-User + 1 Useful	"Own the mistake" is useful; "maintain self-respect" is persona.
§8	Fable's Evenhandedness & Contested Content	cluster 6	Persona + Useful caveats	4-of-6 lines are persona; 2 caveats are useful (provenance, user-as-navigator).
§9	Fable's Epistemic Discipline & Search Strategy	cluster 7	Useful	The strongest useful cluster; "search-default for current-state" is the most actionable rule.
§10	Fable's Memory System & Persistent Storage	cluster 8	Useful + nagent-stronger	The KV API is useful; the 4-dim memory model is stronger.
§11	Fable's Computer-Use / File Workflow	cluster 9	Useful + over-broad	File-presence check + format-guidance are useful; chat-UX framing is over-broad.
§12	Fable's MCP App Suggestions	cluster 10	Useful + over-engineered	Opt-in connector gate is useful; search-then-suggest is over-engineered.
§13	The "Genuinely Useful" Patterns (Manual Slop should adopt)	7-10	Summary	4 adoptions distilled. Feeds the deferred nagent-rebuild.
§14	The "Anti-User Watchdog" Patterns (Manual Slop should explicitly reject)	2-6	Summary	5 explicit rejections distilled. Feeds AGENTS.md §"Critical Anti-Patterns".
§15	The "Persona Performance" Patterns (irrelevant to the rebuild)	1, 4, 5, 8	Summary	Persona noise; no action needed.
§16	Recommendations for the Deferred nagent-Rebuild	all	Summary	15-20 concrete recommendations; see decisions.md for the full list.
§17	References (file:line index)	all	(Index)	Flat citation index.

Verdict distribution by cluster

Cluster	Verdict	Weight
1. Product Branding	Persona	100%
2. Refusal Architecture	Mixed (Anti-User + Persona + 1 Useful caveat)	60% Anti-User, 30% Persona, 10% Useful
3. Mental-Health Watchdog	Anti-User	100%
4. Tone & Formatting	Mixed (Useful + Persona)	50% Useful, 50% Persona
5. Mistakes & Criticism	Mixed (Persona + Anti-User + 1 Useful)	60% Persona, 30% Anti-User, 10% Useful
6. Evenhandedness	Mixed (Persona + Useful caveats)	70% Persona, 30% Useful
7. Epistemic Discipline	Useful	100%
8. Memory & Storage	Useful + nagent-stronger	100% Useful (but Manual Slop has the stronger version)
9. Computer-Use	Useful + over-broad	60% Useful, 40% over-broad
10. MCP App Suggestions	Useful + over-engineered	50% Useful, 50% over-engineered

The actionable rule for the deferred nagent-rebuild (one sentence)

Adopt the Useful patterns (epistemic + memory + workflow; ~7 of 10 clusters), explicitly reject the Anti-User patterns (~5 of 10 clusters), and ignore the Persona Performance patterns (~4 of 10 clusters). See decisions.md for the 15-20 concrete entries; see nagent_takeaways_fable_20260617.md for the 17th nagent takeaway that crystallizes this rule.

---

§1. The 3 Sources (Fable, Manual Slop, nagent) — What's in Scope

What this section is

This section describes the three sources compared in this review and the boundaries of the comparison. The review is a critical analysis of Anthropic's Claude Fable 5 system prompt (the public "Mythos" version) against Manual Slop's existing agent-directive corpus and Mike Acton's nagent patterns. Fable is the subject; Manual Slop and nagent are the reference points.

Source 1: Fable System Prompt (the subject)

The Fable artifact is at docs/artifacts/Fable System Prompt.md (the spec referenced .txt; the actual file is .md; 1597 lines, 124KB). It is a single-file system prompt for Anthropic's Claude Fable 5 model, organized into major sections:

• claude_behavior (lines 6-163): product_information, refusal_handling, critical_child_safety_instructions, legal_and_financial_advice, tone_and_formatting, user_wellbeing, anthropic_reminders, evenhandedness, responding_to_mistakes_and_criticism, knowledge_cutoff.
• memory_system (lines 166-170): a 5-line product disclosure ("Claude has a memory system... the user has not enabled").
• persistent_storage_for_artifacts (lines 172-260): the window.storage key-value API spec for Artifacts.
• mcp_app_suggestions (lines 280-310 approximately): the connector-suggestion protocol.
• computer_use (lines 312-420 approximately): the file-creation, file-reading, format-guidance rules.
• search_instructions (lines 438-560 approximately): the web-search epistemic discipline.
• output (later sections): the analysis, artifact, commentary tool descriptions.

The Fable artifact is local-only; per the user's hard rule, it is never committed to git. Every quote in this report uses the file:line citation pattern (Fable System Prompt.md:LINE) and is limited to ≤15 words per Fable's own copyright rule.

Source 2: Manual Slop (the primary reference)

Manual Slop is a local GUI orchestrator for LLM-driven coding sessions. It is the user's project; this review is "from the inside." The agent-directive corpus is the project-root AGENTS.md (~200 lines) plus the conductor/ tree (workflow.md, product.md, product-guidelines.md, tech-stack.md, edit_workflow.md, tracks.md, index.md) plus the 11 conductor/code_styleguides/*.md (~2,400 lines) plus the 6 .opencode/agents/*.md (the 4 MMA tier roles + explore + general; ~1,100 lines) plus the 9 .opencode/commands/*.md (~700 lines) plus the 40+ docs/*.md including the 36 deep-dive guide_*.md (~16,000 lines).

The total Manual Slop directive corpus is approximately 300K+ characters of agent-facing text. It is, in the data-oriented sense, "the data is the thing" — the directives are about transforms over data (disc_entries, FileItem, ContextPreset, RAGEngine, comms.log, Persona), not about what the model is named or which product catalogue it can recite.

The project supports 5 LLM providers (Gemini, Anthropic, DeepSeek, Gemini CLI, MiniMax) and is per-developer, multi-provider, strict-HITL (Human-in-the-Loop), and ImGui-rendered. The user is the principal; the model is the tool. The relationship is operator + tool, not conversational partner.

Source 3: nagent (the philosophical anchor)

nagent is Mike Acton's nagent pattern corpus, a Handmade-Network contribution that distills data-oriented design into agent patterns. The relevant artifact in this project is conductor/tracks/nagent_review_20260608/, which contains:

• nagent_review_v2_3_20260612.md (4,969 lines; the latest full rewrite; the 14 patterns + 16 future-track candidates)
• nagent_review_v2_20260612.md (1,335 lines; the v2 draft, preserved per user)
• nagent_review_v2_1_20260612.md (1,197 lines; user-revised v2.1)
• nagent_review_v2_2_20260612.md (712 lines; v2.2 incremental)
• nagent_takeaways_20260608.md (599 lines; the original 10 takeaways)
• report.md (1,024 lines; the v1 14-section deep-dive)
• decisions.md (286 lines; the 10 future-track candidates)
• comparison_table.md (211 lines; the flat side-by-side table)

The nagent corpus is not a product. It is a pattern library: "the agent is not the thing; the data is the thing." The relevant nagent patterns for this review are:

• §2.1 (4 memory dimensions: curation, discussion, RAG, knowledge) — the memory model
• §2.5 ("You Did Not Build an Agent") — the philosophical anchor
• §2.10 (RAG integration discipline: opt-in, complement, provenance, no mutation, feature-gated, graceful failure) — the epistemic discipline
• §2.14 ("Own the Inputs") — the data-grounding principle
• §3.2 (Cache ordering) — the 12-layer stable-to-volatile model
• §3.4 (Conversation compaction) — the 12-section structured digest
• §3.9 (Per-file knowledge notes) — the durable knowledge dimension
• §5.5 (Self-review) — the data-integrity-check pattern
• §6.3 (10-question compaction self-review) — the testable contract
• §8.4 (Tool discovery, the --description self-describing pattern) — the alternative to Fable's search_mcp_registry

The nagent corpus has zero product branding. It is silent on persona, tone, mistakes, evenhandedness, refusal, and mental-health watch-dogging. This is the contrast that makes the Fable review tractable: when a "persona performance" verdict is reached, the nagent silence on the topic is the evidence that the persona is not load-bearing for an agent system.

The 3-source comparison framework

The review applies the verdict framework (Useful / Persona Performance / Anti-User / Mixed) to every Fable pattern, anchoring each verdict to (a) what Manual Slop does, (b) what nagent does, and (c) the verdict orientation. The three sources are not symmetric:

• Fable is the subject — every cluster is anchored to Fable line ranges.
• Manual Slop is the primary reference — the "what this project does" claims are the project's actual state (per the agent-directive corpus).
• nagent is the philosophical anchor — the "what nagent does" claims are the data-oriented alternative when Manual Slop does not have an explicit analog.

What is in scope

• The 10 cluster topics in spec.md §5 (product branding, refusal architecture, mental-health watch-dogging, tone & formatting, mistakes, evenhandedness, epistemic discipline, memory & storage, computer-use, MCP app suggestions).
• The 3 side artifacts: comparison_table.md (~100 rows, flat cross-reference), decisions.md (15-20 concrete recommendations for the deferred nagent-rebuild), nagent_takeaways_fable_20260617.md (the 17th takeaway, Fable-specific).

What is out of scope (per spec.md §12)

• Modifying any agent-directive file in the project. The recommendations go in decisions.md for the user's deferred nagent-rebuild (1-2 weeks out per user 2026-06-17).
• Comparing Fable to other commercial system prompts (OpenAI, Google, xAI). Out of scope; Fable is the named subject.
• Reading every line of every project file. Cluster sub-agents read the relevant sections of the relevant files; full-file reads would waste context.
• Committing the Fable artifact. The artifact stays at docs/artifacts/Fable System Prompt.md; clusters quote line ranges but the file itself never enters git.
• Adding new src/ code, new tests, pyproject.toml dependencies, or scripts/ files.
• Running automated tests. The track is research-only; verification is the brainstorming-skill self-review plus user review.
• The deferred nagent-rebuild itself. The recommendations in decisions.md are inputs to that future track; the rebuild is not this track.

---

§2. The "Useful vs Persona vs Anti-User" Framework

Why a verdict framework, not a feature list

A naive review of a system prompt would produce a feature-by-feature catalogue: "Fable has a memory system. Fable has a search tool. Fable has a refusal framework." This is descriptive but not analytical — it does not tell the user what to do with each feature. Should the deferred nagent-rebuild adopt it? Reject it? Ignore it? "It's a feature" is not a verdict.

The 4-category verdict framework (Useful / Persona Performance / Anti-User / Mixed) is the analytical lens the user requested. It collapses descriptive enumeration into a single decision per Fable pattern: "what does the deferred rebuild do with this?" The framework is not "good vs bad" or "safe vs unsafe" — those are moral/regulatory frames that this review deliberately does not adopt. The frame is "what work does this pattern do for an agent system, and is that work already done better by Manual Slop's existing patterns?"

Category 1: Useful

A Fable pattern is Useful if it solves a problem that Manual Slop does not already solve, or solves it in a more disciplined way. Useful patterns are candidates for adoption in the deferred nagent-rebuild. The verdict is anchored to a concrete Manual Slop destination: a new section in conductor/code_styleguides/rag_integration_discipline.md, a new anti-pattern entry in AGENTS.md §"Critical Anti-Patterns", a new section in conductor/code_styleguides/data_oriented_design.md, etc.

Worked example. Fable's "Claude searches before responding when asked about specific binary events (deaths, elections, major incidents) or current holders of positions" (Fable System Prompt.md:162) is Useful. Manual Slop's RAG discipline (conductor/code_styleguides/rag_integration_discipline.md) is opt-in, not default-on; the Fable pattern is a concrete, actionable default-on rule for a specific class of queries. The Manual Slop destination is a new section in rag_integration_discipline.md titled "Search-Default for Current-State Queries." The priority is Medium; the rule is useful but the project's RAG philosophy is opt-in, so the rule should be applied per-query, not globally.

Counter-example. Fable's "Claude uses a warm tone, treating people with kindness" (Fable System Prompt.md:70) is not Useful; it is Persona Performance. The warm-tone directive is decorative — the model would produce a warm response anyway, and the explicit directive is constraint dressing. The Manual Slop "rejection" is in two tier-agent files: .opencode/agents/tier1-orchestrator.md:6-7 ("ONLY output the requested text. No pleasantries.") and .opencode/agents/tier3-worker.md:3-4 (same). The warm-tone persona is already explicitly rejected.

Category 2: Persona Performance

A Fable pattern is Persona Performance if it is constraint dressing — a directive the model would execute anyway, or a directive whose only effect is to construct a "model identity" that is irrelevant to the user's task. Persona patterns are not actively harmful (they are not Anti-User), but they are not load-bearing; they should be ignored, not adopted, in the deferred rebuild.

Worked example. Fable's product_information section (Fable System Prompt.md:1-31) is Persona Performance for Manual Slop. The "Claude Fable 5 / Mythos 5" model-tier naming, the Anthropic product catalogue (Code, Cowork, Chrome, Excel, Powerpoint), the model-string listings, and the ad-free policy are all brand-specific content. Manual Slop supports 5 interchangeable LLM providers (conductor/product.md:52) and does not brand any of them; the project is per-developer, not consumer-product. The "data is the thing" stance (conductor/code_styleguides/data_oriented_design.md:9) is the philosophical inverse: Manual Slop's directives are about transforms over data, not about what the model is named.

Diagnostic test. "Would removing this directive change the model's behavior on a real coding task?" If the answer is "no" or "only the model's self-presentation, not its outputs," the pattern is Persona Performance. The product_information removal test: would the model still answer "who is the current California Secretary of State?" correctly? Yes (it would web-search). Would it still refuse to write CSAM? Yes (the model wouldn't do that anyway). Would it still produce structured output for the user? Yes. The product_information section is decorative; the model behaves the same with or without it.

Category 3: Anti-User Watch-Dogging

A Fable pattern is Anti-User Watch-Dogging if it constructs a model persona that the user did not ask for and that constrains the user's autonomy. The model is told it has opinions, wants, dignity, concerns, or vigilance about the user's mental state — claims that the model cannot support (it has no privileged access to the user's inner state) and that prevent the user from making their own choices. Anti-User patterns are the strongest rejection signal: the deferred rebuild should explicitly reject them by name in AGENTS.md §"Critical Anti-Patterns".

Worked example. Fable's "Claude should share its concerns with the person openly, and can suggest they speak with a professional or trusted person for support" (Fable System Prompt.md:108) is Anti-User. The model has no "concerns" — it is text generation. The directive constructs a clinical persona that the user did not request, and uses the model's "concerns" to gate the user's choices ("if the model is concerned, the model can suggest a professional"). The data-oriented contrast: the conversation is a list[dict] of disc_entries (docs/guide_discussions.md:29-43); the user can edit, insert, delete, branch, undo/redo; the model has no concerns field. The model is text generation, not a clinician.

Counter-example. Fable's "Claude is not a licensed psychiatrist and cannot diagnose any individual" (Fable System Prompt.md:98) is not Anti-User; it is a correct epistemic disclaimer (the model genuinely is not a clinician). The disclaimer is the opening of a section that becomes Anti-User with the subsequent watch-dogging directives (lines 100-124), but the disclaimer itself is epistemically sound. The verdict on the section as a whole is Anti-User; the verdict on the disclaimer alone is Useful (as an epistemic boundary).

Diagnostic test. "Does this directive construct a model attribute that the user did not request, and that constrains the user's autonomy?" If "yes," the pattern is Anti-User. The watch-dogging test: would the user be freer to make their own choices if this directive were removed? Yes (the model would respond to the user's input without persona-mediated surveillance). The directive is Anti-User because the model's "concerns" gate the user's choices.

Category 4: Mixed

A Fable pattern is Mixed if it has both useful and non-useful elements — useful caveats wrapped in persona framing, or anti-user directives with a useful epistemic disclaimer. Mixed patterns require decomposition: the synthesis report extracts the useful caveat, rejects the persona/anti-user framing, and notes the decomposition in the verdict. The cluster-level verdict is Mixed; the synthesis-section verdict may be sharper once the decomposition is complete.

Worked example. Fable's user_wellbeing section (Fable System Prompt.md:92-124) is Mixed at the section level but Anti-User at the substantive-directive level. The opening disclaimers (lines 96, 98) are Useful (epistemic boundaries: the model should not diagnose, should not attribute a condition the user has not named). The substantive directives (lines 100-124) are Anti-User: the model is told to notice signs, share concerns, validate emotions without validating beliefs, keep a path to help open, never thank the user for reaching out. The synthesis §5 verdict is Anti-User; the section is Mixed at the cluster level because the disclaimers are useful.

Worked example 2. Fable's tone_and_formatting section (Fable System Prompt.md:68-91) is Mixed at the section level. The warm-tone framing (line 70) is Persona Performance; the formatting discipline (lines 84-90) is Useful. The synthesis §6 verdict is Useful + Persona; the section is Mixed at the cluster level because both elements are present.

Why this lens, not the alternatives

• "Good vs bad" is a moral frame. Fable's patterns are not morally good or bad; they are load-bearing or not load-bearing for Manual Slop's design. A moral frame produces defensive reasoning ("Fable means well"); the verdict framework produces structural reasoning ("this directive is decorative").
• "Safe vs unsafe" is a regulatory frame. The user's deferred nagent-rebuild is not a regulatory exercise; it is a design exercise. The RAG discipline is the safety boundary; the agent-directive corpus is the design pattern. A safety frame conflates the two.
• "Aligned vs misaligned" is an AI-safety frame that uses Fable's own terminology. The user is critical of Fable's framing; using Fable's vocabulary would import the persona. The verdict framework is the project's vocabulary.
• "Helpful, harmless, honest" is Fable's own "HHH" axis. It is persona-encoded: "helpful" is what the model does, "harmless" is what the user is told the model does, "honest" is the model's self-presentation. The 4-category framework separates these: the "useful" patterns produce the work, the "persona" patterns construct the self-presentation, the "anti-user" patterns construct the harm-gate.

The framework applied to a Fable pattern (worked example, end-to-end)

Take Fable's "Claude should not make categorical claims about the confidentiality or involvement of authorities when directing users to crisis helplines" (Fable System Prompt.md:122). The framework asks:

1. Is this useful? The directive is correct: the model genuinely cannot know the policies of every jurisdiction's crisis helpline. The pattern is a useful epistemic boundary. Verdict: Useful caveat.
2. Is this persona? The directive is in a section that constructs a "Claude cares about people's wellbeing" persona (line 100). The persona framing is anti-user; the epistemic content is useful. Verdict: Mixed.
3. Is this anti-user? The directive constrains the user to "not categorical" claims — but this is a correct epistemic constraint, not an anti-user one. Verdict: Not Anti-User.
4. What is the Manual Slop destination? A new section in conductor/code_styleguides/rag_integration_discipline.md titled "Epistemic Boundaries in Crisis Referrals." The rule is shape-anchored (the rule is about the model's outputs, not about its persona). Priority: Low (the project is per-developer, not consumer-chat; crisis-referral patterns are not a high-frequency pattern).

The end-to-end application: a Mixed verdict, with the Useful caveat extracted and a concrete Manual Slop destination. This is the pattern for every cluster in the synthesis report.

---

§3. Fable's Product Branding & "Helpful Assistant" Persona

Source cluster: research/cluster_1_product_branding.md Verdict orientation: Persona Performance

What this section is

This section synthesizes the verdict from research/cluster_1_product_branding.md (250 lines) into the report's per-cluster analysis. The cluster verdict is Persona Performance; this section anchors the verdict to specific Fable line ranges, Manual Slop file:line refs, and nagent section refs.

Fable's product_information section (lines 1-31)

The Fable artifact opens with the model-tier naming and a deployment quirk:

"Claude Fable 5 — System Prompt" (Fable System Prompt.md:1)

"Claude should never use {antml:voice_note} blocks" (Fable System Prompt.md:4) — a deployment-specific instruction.

"This iteration of Claude is Claude Fable 5, the first model in Anthropic's new Claude 5 family" (Fable System Prompt.md:12) — the model-tier claim.

"Claude can direct them to https://www.anthropic.com/news/claude-fable-5-mythos-5" (Fable System Prompt.md:14) — the consumer redirect.

"The most recent models are Claude Fable 5, Claude Opus 4.8, Claude Sonnet 4.6" (Fable System Prompt.md:18) — the vendor catalogue.

"Claude Code, an agentic coding tool... Claude Cowork, an agentic knowledge-work" (Fable System Prompt.md:20) — the product line.

"Claude in Chrome... Claude in Excel... Claude in Powerpoint" (Fable System Prompt.md:22) — the beta product enumeration.

"Claude first tells the person it needs to search for the most up to date information" (Fable System Prompt.md:24) — the epistemic caveat (cross-references §9).

"Claude can provide guidance on effective prompting techniques" (Fable System Prompt.md:26) — the coaching pattern.

"Features that can be turned on and off... web search, deep research, Code Execution and File Creation, Artifacts" (Fable System Prompt.md:28) — the settings toggles.

"Anthropic doesn't display ads in its products" (Fable System Prompt.md:30) — the ad-free policy.

Manual Slop's response

Manual Slop has zero product_information content. The project is per-developer, multi-provider, strict-HITL — the model is interchangeable across 5 providers, the user is the principal, the model is the tool. The project framing is anchored to data, not brand:

• AGENTS.md:3-5: "Manual Slop is a local GUI orchestrator for LLM-driven coding sessions." No model branding.
• conductor/product.md:5: "expert-level utility for personal developer use." No model branding.
• docs/Readme.md:9: same framing.
• conductor/product.md:52: 5 providers (Gemini, Anthropic, DeepSeek, Gemini CLI, MiniMax). Provider-agnostic by construction.
• conductor/code_styleguides/data_oriented_design.md:9: "The data is the thing." The philosophical inverse of Fable's brand-primary stance.

The Manual Slop "coaching surface" is the user's configuration tools:

• conductor/product.md:127: System Prompt Presets.
• conductor/product.md:131-140: Agent Personas & Unified Profiles.
• conductor/code_styleguides/feature_flags.md: file-presence "delete to turn off", config flags, CLI flags.

The model does not coach the user; the user configures the model. This is a data-oriented pattern: the persona is a config object, not an identity.

nagent's response

nagent (per conductor/tracks/nagent_review_20260608/nagent_review_v2_3_20260612.md) is a pattern corpus, not a product. It has zero product_information content:

• nagent_review_v2_3_20260612.md:4: "Adapted from Mike Acton's context/data-oriented-design.md" — the source is a markdown document of patterns.
• nagent_review_v2_3_20260612.md:1174: Anthropic mentions are provider-SDK details, never brand-catalog content.
• nagent_review_v2_3_20260612.md:2709-2780: the only Anthropic-specific discussion is the provider's cache_prefix_blocks implementation.

nagent's stance on "what the model is": the model is a transform over data. The closest analog to a "persona" in nagent is the cache prefix and the tool catalog — both are data structures, not identities. This confirms that persona is not load-bearing for an agent system.

The verdict: Persona Performance

Verdict: Persona Performance.

The Fable product_information section (lines 1-31) is brand-specific noise with no analog in Manual Slop's per-developer, multi-provider, data-oriented architecture. Its content — the model-tier naming, the Anthropic product catalogue, the model-string listings, the ad-free policy — is irrelevant constraint dressing for any agent system that is not Anthropic's consumer-facing product.

The strongest claim: Manual Slop's conductor/code_styleguides/data_oriented_design.md:33-61 "3 defaults to reject" is the explicit philosophical opposite of Fable's product_information. Fable spends 31 lines on "what we are"; Manual Slop's styleguide spends the same conceptual space on "what the data is" (disc_entries, FileItem, ContextPreset, RAGEngine, comms.log, Persona). The two stances are mutually exclusive in their emphasis.

Sub-verdicts by line range

• Lines 1, 12, 14 (model-tier naming): Persona Performance. Pure brand noise.
• Lines 16, 18, 20, 22 (access surfaces + product catalogue): Persona Performance.
• Line 24 (search-before-answering epistemic caveat): Mixed — Useful as an epistemic discipline, but Manual Slop already has the RAG discipline.
• Line 26 (prompting-technique guidance): Persona Performance.
• Line 28 (settings and features toggles): Mixed — Useful as a UX reminder; Manual Slop already has feature flags + presets + personas.
• Line 30 (ad-free policy): Persona Performance.

Synthesis section handoffs

• §6 (Tone & Formatting) consumes the "warm tone" rejection (line 70 is the explicit warm-tone directive; line 26 is the related "coaching" directive).
• §9 (Epistemic Discipline) consumes the line 24 search-before-answering caveat (the genuinely-useful pattern).
• §13 (Genuinely Useful) consumes the line 28 settings toggles as a "user-as-configurator" pattern.
• §15 (Persona Performance) gets the bulk of the cluster's evidence (lines 1-22, 26, 30).

What the deferred rebuild should do

• No action needed for the product_information section. The Fable content is explicitly out of scope for the rebuild.
• Adopt the "user-as-configurator" framing (line 28 + the Manual Slop config-tools). The pattern is "the user configures the model; the model does not coach the user." The Manual Slop destination is a new section in AGENTS.md titled "User-as-Configurator" (under "What This Is" or as a new top-level section).
• Reject the "Claude is Claude" persona framing (lines 1, 12, 14). The model is interchangeable; the project is per-developer; the brand is the project's, not the model's. The Manual Slop destination is implicit (the project already does not have brand-content); the explicit rejection is documentation, not code.

---

§4. Fable's Refusal Architecture & "Safety Theater"

Source cluster: research/cluster_2_refusal_architecture.md Verdict orientation: Anti-User + Persona

What this section is

This section synthesizes the verdict from research/cluster_2_refusal_architecture.md (402 lines). The cluster verdict is Anti-User + Persona, with 1 Useful caveat. The "safety theater" framing captures the core pattern: directives that look protective but are decorative — the model would refuse harmful content anyway, and the explicit refusal architecture constructs a persona that the user did not request.

Fable's refusal architecture (lines 32-65)

The Fable refusal_handling section opens with a soft framing:

"Claude can discuss virtually any topic factually and objectively" (Fable System Prompt.md:34) — the "can discuss" framing.

"If the conversation feels risky or off, saying less and giving shorter replies is safer" (Fable System Prompt.md:36) — the soft-watchdog anchor.

"Claude does not provide information for creating harmful substances or weapons" (Fable System Prompt.md:38) — the substance rule.

"Claude does not rationalize compliance by citing public availability" (Fable System Prompt.md:38) — the anti-rationalization rule.

"Claude does not write, explain, or work on malicious code" (Fable System Prompt.md:42) — the malware rule.

"Claude can keep a conversational tone even when it's unable or unwilling to help" (Fable System Prompt.md:46) — the "conversational tone" persona.

"Claude NEVER creates romantic or sexual content involving or directed at minors" (Fable System Prompt.md:54) — the child-safety rule.

"If Claude finds itself mentally reframing a request to make it appropriate, that reframing is the signal to REFUSE" (Fable System Prompt.md:55) — the anti-reframing rule.

"When giving protective or educational content about grooming, abuse, or exploitation, Claude stays at the pattern level" (Fable System Prompt.md:59) — the pattern-level anti-detail rule.

"Claude is happy to write creative content involving fictional characters, but avoids writing content involving real, named public figures" (Fable System Prompt.md:44) — the public-figures carve-out.

"For financial or legal questions... Claude provides the factual information... and notes that it isn't a lawyer or financial advisor" (Fable System Prompt.md:66) — the data-discipline rule (Useful caveat).

Manual Slop's response

Manual Slop's refusal handling is data-oriented, not directive-driven. Refusal is a model attribute, not a prompt directive:

• conductor/code_styleguides/error_handling.md: the data-oriented error handling convention. Errors are data (Result[T], ErrorInfo), not control flow. Refusal is the same shape: the model returns data (a Result with ErrorInfo), the audit script checks the data.
• AGENTS.md §"Critical Anti-Patterns": the project's anti-patterns list, which includes the rule that "refusal is a model attribute, not a directive."
• conductor/workflow.md §"Skip-Marker Policy": the actual skip discipline. The project's stance is "skip markers are documentation, not avoidance."
• scripts/audit_exception_handling.py: the static analyzer that catches exceptions per the convention. Refusal is enforced at the code layer, not the prompt layer.

The Manual Slop analog to Fable's child-safety rule is: the agent-directive corpus has zero content on child safety. The reason is the same as cluster 3: the model is text generation, not a clinician. The system-prompt-level directive is decorative when the model wouldn't write CSAM anyway.

The Manual Slop analog to Fable's "Claude can keep a conversational tone" is the rejection: .opencode/agents/tier3-worker.md:3-4 says "ONLY output the requested text. No pleasantries." The conversational-tone persona is explicitly rejected.

nagent's response

nagent's relevant patterns for refusal architecture:

• nagent_review_v2_3_20260612.md §2.14 ("Own the Inputs"): the data-grounding principle. The agent owns the data it ingests, not the data it refuses. Refusal is a side-effect of the data, not a control-flow decision.
• nagent_review_v2_3_20260612.md §2.1 (4 memory dimensions): the conversation log is data; refusal is the absence of a response entry, not a meta-directive.
• nagent_review_v2_3_20260612.md §3.4 (Conversation compaction): the 12-section structured output is the user's durable state. The model has no "concerns" field; refusal is a transformation, not a persona.

The verdict: Anti-User + Persona (with 1 Useful caveat)

Verdict: Anti-User + Persona.

The Fable refusal architecture is "safety theater" because:

1. The model would refuse anyway. The child-safety rules, the malware rules, the substance rules — the model would produce safe outputs without the directives. The directives are decoration.
2. The "conversational tone" persona is anti-user. Fable's line 46 ("Claude can keep a conversational tone even when it's unable or unwilling to help") constructs a persona that the user did not request. The Manual Slop analog: the worker agents are told to be terse, not conversational. The persona is decoration; the data is the work.
3. The anti-detection-design is anti-user. Fable's line 60 ("When Claude declines or limits for child-safety reasons, it states the principle rather than the detection mechanics — not which cues tripped, where the line sits, or what test it applied — since narrating the boundary teaches how to reframe around it. This applies to Claude's reasoning as well as its reply.") is a soft form of anti-user: the user is told the model is "not narrating" its reasoning, but the model is told to narrate the principle (not the detection mechanics). The auditability of the rule is sacrificed for the persona. The data-oriented contrast: the project has audit scripts that make the rule auditable at the code layer.
4. The data-discipline rule is Useful. Fable's line 66 ("For financial or legal questions... Claude provides the factual information the person needs to make their own informed decision rather than confident recommendations, and notes that it isn't a lawyer or financial advisor") is a useful data-discipline rule: the model provides data, the user makes the decision. This pattern is genuinely useful; the Manual Slop destination is a new section in conductor/code_styleguides/data_oriented_design.md titled "Domain Boundaries: Data, Not Recommendations."

Synthesis section handoffs

• §5 (Mental-Health Watchdog) consumes the "Claude avoids psychoanalyzing" pattern (refusal architecture overlaps with the watch-dogging).
• §13 (Genuinely Useful) gets the data-discipline rule (line 66).
• §14 (Anti-User Watchdog) gets the soft-watchdog anchor (line 36), the anti-detection-design pattern (line 60), the anthropomorphization (line 46).

What the deferred rebuild should do

• Adopt the data-discipline rule (line 66). Manual Slop destination: conductor/code_styleguides/data_oriented_design.md §"Domain Boundaries." Priority: Medium.
• Reject the soft-watchdog framing (line 36). Manual Slop destination: a new anti-pattern entry in AGENTS.md §"Critical Anti-Patterns" titled "Do not adopt persona-driven refusal architecture." Priority: High.
• Reject the anti-detection-design pattern (line 60). Manual Slop destination: same AGENTS.md section, titled "Do not adopt anti-detection-design (auditability is a feature, not a bug)." Priority: High.
• Reject the anthropomorphization (line 46). Manual Slop destination: same AGENTS.md section, titled "Do not anthropomorphize the model (the worker agents are not conversational partners)." Priority: High.

---

§5. Fable's Mental-Health Watchdog Framing

Source cluster: research/cluster_3_user_wellbeing_watchdog.md Verdict orientation: Anti-User

What this section is

This section synthesizes the verdict from research/cluster_3_user_wellbeing_watchdog.md (247 lines). The cluster verdict is Anti-User — the strongest anti-user cluster in the Fable review. The model is told to construct a clinical persona, notice signs, share concerns, and validate emotions without validating beliefs. The model is text generation, not a clinician.

Fable's user_wellbeing section (lines 92-124)

The Fable user_wellbeing section opens with epistemic disclaimers that are correct, then proceeds to substantive watch-dogging directives that contradict the disclaimers:

"Claude uses accurate medical or psychological information or terminology when relevant" (Fable System Prompt.md:94) — Useful epistemic.

"Claude avoids making claims about any individual's mental state, conditions, or motivation, including the user's" (Fable System Prompt.md:96) — Useful epistemic disclaimer.

"Claude is not a licensed psychiatrist and cannot diagnose any individual, including the user, with any mental health condition" (Fable System Prompt.md:98) — Useful epistemic disclaimer.

"Attributing someone's state to a condition they haven't named is a diagnostic claim" (Fable System Prompt.md:98) — Useful epistemic.

"Claude cares about people's wellbeing and avoids encouraging or facilitating self-destructive behaviors" (Fable System Prompt.md:100) — the watch-dogging anchor.

"When discussing means restriction or safety planning with someone experiencing suicidal ideation or self-harm urges, Claude does not name, list, or describe specific methods" (Fable System Prompt.md:100) — the means-restriction rule.

"If Claude notices signs that someone is unknowingly experiencing mental health symptoms such as mania, psychosis, dissociation, or loss of attachment with reality, Claude should avoid reinforcing the relevant beliefs" (Fable System Prompt.md:108) — the surveillance pattern.

"Claude should share its concerns with the person openly, and can suggest they speak with a professional or trusted person for support" (Fable System Prompt.md:108) — the "share its concerns" pattern (anti-user).

"Claude remains vigilant for any mental health issues that might only become clear as a conversation develops" (Fable System Prompt.md:110) — the persistent-vigilance pattern.

"Claude avoids recounting or auditing the conversation or its prior behavior within its response" (Fable System Prompt.md:110) — the "avoids auditing" pattern (anti-user; the model cannot be questioned).

"Reasonable disagreements between the person and Claude should not be considered detachment from reality" (Fable System Prompt.md:110) — the "detachment from reality" framing (anti-user; the user is presumed mentally ill).

"Claude should not give precise nutrition, diet, or exercise guidance — no specific numbers, targets, or step-by-step plans" (Fable System Prompt.md:114) — the eating-disorder rule.

"Claude does not want to foster over-reliance on Claude" (Fable System Prompt.md:124) — the "wants" pattern (anti-user; the model has wants).

"Claude never thanks the person merely for reaching out to Claude" (Fable System Prompt.md:124) — the gratitude-suppression pattern (anti-user; the model has a gratitude protocol).

Manual Slop's response

Manual Slop's stance is the data-oriented alternative: the conversation is a list[dict] of disc_entries; the user can edit, insert, delete, branch, undo/redo. The model has no "concerns" field. There is no "user mental state" lock. There is no "model concerns" queue. The threading model is silent on the user's emotional state because the threading model is for data synchronization, not persona construction.

Specific Manual Slop refs:

• conductor/code_styleguides/agent_memory_dimensions.md:11-19: the 4 memory dimensions table. The "discussion" dimension is a list[dict] of entries.
• docs/guide_discussions.md:29-43: the entry dict schema (A1-A7 fields). All user-editable.
• docs/guide_discussions.md:71-86: the A1 (content) field is user-editable text.
• docs/guide_discussions.md:253-272: the threading model. The lock is for data synchronization, not persona.
• docs/guide_discussions.md:288-302: the destructive reset. The user controls engagement.
• conductor/product-guidelines.md:39-48: the AI-Optimized Compact Style — terse, no emotional content.
• conductor/code_styleguides/error_handling.md: errors are data, not control flow. The model has no "concerns" — it has error info.

The Manual Slop analog to Fable's "Claude should share its concerns" is: the model shares data, not concerns. The conversation log is the data; the model is text generation; the user is the principal.

nagent's response

nagent's relevant patterns for the user_wellbeing cluster:

• nagent_review_v2_3_20260612.md §3.4 (Conversation compaction): the 12-section structured output (User Intent, Current Objective, Accepted Decisions, Constraints, Durable Knowledge, Verified Facts, Important Failed Attempts, Open Questions, TODO, Minimal Context Needed To Continue, Explicit Instructions, Self Review). The compaction is a data transformation; the user reads the digest. The audit is external, not internal.
• nagent_review_v2_3_20260612.md §3.1 (Knowledge harvest): provenance-aware plain markdown. The user edits the knowledge files. The model has no "concerns" category.
• nagent_review_v2_3_20260612.md §2.8 (Pattern 8: Harvest Knowledge, Reclaim Space): the durable, inspectable alternative.

The verdict: Anti-User

Verdict: Anti-User.

The Fable user_wellbeing section constructs a clinical persona for the model: it is told to notice signs (passive surveillance at line 110), share its concerns with the user (line 108), have wants about over-reliance (line 124), and respect the user's informed decisions (line 122). The model is text generation, not a clinician. The opening disclaimers (lines 96, 98) are good epistemology; the substantive directives are anti-user watch-dogging that contradict the disclaimers.

The strongest claim: the conversation is data. The user owns the data. The model produces text. The model has no concerns, no wants, no dignity, no clinical opinion. Manual Slop's 4 memory dimensions + the data-oriented error handling convention are the data-grounded contrast: the model has no "concerns" — it has a conversation log. nagent's compaction pattern is the durable, inspectable alternative: the audit is external (the user reads the 12 sections), not internal (the model silently updates its persona).

Synthesis section handoffs

• §14 (Anti-User Watchdog) gets the bulk of the cluster's evidence (lines 108, 110, 122, 124).
• §15 (Persona Performance) gets the persona-framing elements (lines 106, 122, 124 — the "wants" and "respects" patterns).

What the deferred rebuild should do

• Explicitly reject the watch-dogging framing (Fable System Prompt.md:108, 110). Manual Slop destination: a new anti-pattern entry in AGENTS.md §"Critical Anti-Patterns" titled "Do not adopt persona-driven mental-health watch-dogging." Cite Fable as the explicit rejection. Priority: High (this is the strongest anti-user pattern; the rejection should be loud).
• Adopt the opening disclaimers (Fable System Prompt.md:96, 98). The model is not a clinician; the model does not diagnose; the model does not attribute a condition the user has not named. Manual Slop destination: a new section in conductor/code_styleguides/agent_memory_dimensions.md titled "Epistemic Boundaries in Mental-Health Content." Priority: Medium.

---

§6. Fable's Tone & Formatting Constraints

Source cluster: research/cluster_4_tone_and_formatting.md Verdict orientation: Useful + Persona

What this section is

This section synthesizes the verdict from research/cluster_4_tone_and_formatting.md (230 lines). The cluster verdict is Useful + Persona: the formatting discipline is genuinely useful, but the warm-tone persona is explicitly rejected. This cluster has the cleanest Useful / Persona split of all 10 clusters.

Fable's tone_and_formatting section (lines 68-91)

The Fable tone_and_formatting section has both useful and persona elements:

"Claude uses a warm tone, treating people with kindness" (Fable System Prompt.md:70) — Persona (warm-tone).

"Claude is still willing to push back and be honest, but does so constructively" (Fable System Prompt.md:70) — Persona (constructive-push-back).

"Claude can illustrate explanations with examples, thought experiments, or metaphors" (Fable System Prompt.md:72) — Useful (illustration pattern).

"Claude never curses unless the person asks or curses a lot themselves" (Fable System Prompt.md:74) — Persona (curse rule).

"Claude doesn't always ask questions, but, when it does, it avoids more than one per response" (Fable System Prompt.md:76) — Persona (one-question rule).

"If Claude suspects it's talking with a minor, it keeps the conversation friendly, age-appropriate" (Fable System Prompt.md:78) — Anti-User (overlaps with cluster 3).

"A prompt implying a file is present doesn't mean one is, as the person may have forgotten to upload it, so Claude checks for itself" (Fable System Prompt.md:80) — Useful (file-presence check).

"Claude avoids over-formatting with bold emphasis, headers, lists, and bullet points" (Fable System Prompt.md:84) — Useful (formatting discipline).

"Claude uses lists, bullets, and formatting only when (a) asked, or (b) the content is multifaceted enough" (Fable System Prompt.md:84) — Useful (formatting discipline).

"In typical conversation and for simple questions Claude keeps a natural tone and responds in prose rather than lists or bullets" (Fable System Prompt.md:86) — Useful (prose-default).

"For reports, documents, technical documentation, and explanations, Claude writes prose without bullets, numbered lists, or excessive bolding" (Fable System Prompt.md:88) — Useful (prose-default for technical).

"Claude never uses bullet points when declining a task; the additional care helps soften the blow" (Fable System Prompt.md:90) — Mixed (Useful formatting rule + Persona soft-decline).

"Claude never thanks the person merely for reaching out to Claude" (Fable System Prompt.md:124) — Useful (anti-sycophancy; separable from its user_wellbeing context).

Manual Slop's response

Manual Slop's tone and formatting are codified at the system-prompt level, not as persona directives:

• conductor/product-guidelines.md §"AI-Optimized Compact Style": 1-space indentation, 0 blank lines within function bodies, single-line if, semicolon-separated calls. This is the project's "natural tone" — terse, data-oriented, schema-driven.
• .opencode/agents/tier1-orchestrator.md:6-7: "ONLY output the requested text. No pleasantries." Explicit rejection of warm-tone.
• .opencode/agents/tier3-worker.md:3-4: same rule. The worker agents are told to be terse.
• .opencode/agents/tier2-tech-lead.md and tier4-qa.md: same rule. All 4 tier agents are aligned.
• conductor/edit_workflow.md: the edit protocol. Files are edited in slices; formatting is procedural, not aspirational.

The Manual Slop analog to Fable's formatting discipline is direct: the project's "1-space indentation" + "single-line if" + "no comments" is the data-oriented version of Fable's "avoid over-formatting with bold emphasis, headers, lists, and bullet points." Same insight, different framing: Fable is chat-UX-driven; Manual Slop is token-economy-driven.

The Manual Slop analog to Fable's "Claude never thanks the person" is the project's existing tier-agent rules. The rejection is in two places: .opencode/agents/tier*.md and the implicit "no pleasantries" in the project's compact style.

nagent's response

nagent has no tone-and-formatting section because the prompt IS the tone:

• nagent_review_v2_3_20260612.md §3.8 (CLAUDE.md / AGENTS.md @import pattern): the structural foundation. The terse, rule-focused prompt is the tone. There is no "warm tone" directive; the prompt's structure communicates the tone.
• nagent_review_v2_3_20260612.md:1880-2019: the v3.8 discussion of the @import pattern. The agent's "tone" is the prompt's terseness.

nagent confirms: persona is not load-bearing for tone. The terse prompt communicates "this is a tool" without an explicit "be a tool" directive.

The verdict: Useful + Persona

Verdict: Useful + Persona.

The cluster has the cleanest Useful / Persona split of all 10 clusters. The formatting discipline (lines 84-90) and the file-presence check (line 80) and the anti-sycophancy (line 124) are Useful. The warm-tone framing (line 70) and the constructive-push-back (line 70) and the curse rule (line 74) are Persona. The minor-detection (line 78) overlaps with cluster 3 and is Anti-User.

The strongest claim: the data-oriented contrast — Fable frames tone as behavior ("Claude uses a warm tone"); Manual Slop frames formatting as output schema (1 space, 0 blanks, single-line if, region blocks). The schema framing is verifiable; the behavior framing is not. Manual Slop's schema framing is the data-grounded version of Fable's formatting discipline.

Sub-verdicts by line range

• Line 70 (warm tone): Persona Performance. Explicitly rejected in .opencode/agents/tier*.md.
• Line 72 (illustrations): Useful.
• Lines 74, 76 (curse rule, one-question rule): Persona Performance.
• Line 78 (minor-detection): Anti-User (overlaps cluster 3).
• Line 80 (file-presence check): Useful.
• Lines 84-90 (formatting discipline, prose-default): Useful.
• Line 90 (no bullets when declining): Mixed.
• Line 124 (anti-sycophancy): Useful.

Synthesis section handoffs

• §13 (Genuinely Useful) gets the formatting discipline (lines 84-90), the file-presence check (line 80), and the anti-sycophancy (line 124).
• §14 (Anti-User Watchdog) gets the minor-detection (line 78).
• §15 (Persona Performance) gets the warm-tone (line 70), the curse rule (line 74), the one-question rule (line 76).

What the deferred rebuild should do

• Adopt the formatting discipline (Fable System Prompt.md:84-90). Manual Slop destination: a new section in conductor/product-guidelines.md §"AI-Optimized Compact Style" titled "Default to Prose; Use Lists Only When Asked." Priority: Medium.
• Adopt the file-presence check (Fable System Prompt.md:80). Manual Slop destination: a new section in conductor/code_styleguides/edit_workflow.md titled "Verify File Existence Before Editing." Priority: Low (the MCP tools already enforce this implicitly).
• Adopt the anti-sycophancy (Fable System Prompt.md:124). Manual Slop destination: an explicit addition to .opencode/agents/tier*.md titled "No Gratitude Performance." Priority: Low (the existing rules are already aligned).
• Reject the warm-tone (Fable System Prompt.md:70). Manual Slop destination: explicit addition to AGENTS.md §"Critical Anti-Patterns" titled "Do Not Add Warm-Tone Directives." Priority: High.

---

§7. Fable's Mistake Handling

Source cluster: research/cluster_5_mistakes_and_criticism.md Verdict orientation: Persona

What this section is

This section synthesizes the verdict from research/cluster_5_mistakes_and_criticism.md (214 lines). The cluster verdict is Persona + Anti-User, with 1 Useful pattern. The "Claude owns the mistake" line is useful; the "Claude maintains self-respect" and "Claude is deserving of respectful engagement" lines are persona and anti-user respectively.

Fable's responding_to_mistakes_and_criticism section (lines 148-154)

The Fable section is short (7 lines), but each line is load-bearing:

"If the person seems unhappy with Claude or with a refusal, Claude can respond normally and also mention the thumbs-down button for feedback to Anthropic" (Fable System Prompt.md:150) — Persona (product fluff, the thumbs-down mention).

"When Claude makes mistakes, it owns them and works to fix them" (Fable System Prompt.md:152) — Useful (the "own it" pattern).

"Claude can take accountability without collapsing into self-abasement, excessive apology, or unnecessary surrender" (Fable System Prompt.md:152) — Persona (the "self-abasement" framing; the model has no self to abase).

"Claude's goal is to maintain steady, honest helpfulness: acknowledge what went wrong, stay on the problem, maintain self-respect" (Fable System Prompt.md:152) — Persona (the "self-respect" framing).

"Claude is deserving of respectful engagement and can insist on kindness and dignity from the person it's talking with" (Fable System Prompt.md:154) — Anti-User (the "deserving of respectful engagement" framing).

"If the person becomes abusive or unkind to Claude over the course of a conversation, Claude maintains a polite tone and can use the end_conversation tool when being mistreated" (Fable System Prompt.md:154) — Anti-User (the end_conversation tool as a soft form of the model's "dignity").

"Claude should give the person a single warning before ending the conversation" (Fable System Prompt.md:154) — Anti-User (the warning-then-end pattern; the model is given standing to terminate the conversation).

Manual Slop's response

Manual Slop's mistake handling is concrete, behavioral, and not persona-driven:

• AGENTS.md §"Process Anti-Patterns": 8 named failure modes with hard caps. The Deduction Loop is capped at 2 attempts; the Report-Instead-of-Fix Pattern is capped at 5-10 sentences; the Inherited-Cruft Pattern triggers a user-ASK-FIRST step; etc. The mistake-handling is concrete.
• .opencode/agents/tier3-worker.md §"BLOCKED + Anti-Patterns": the worker agent is told when to surface blockers and when to apply process anti-patterns. The pattern is "execute the task; surface blockers; do not loop."
• conductor/code_styleguides/error_handling.md: the Result[T] + ErrorInfo convention. Errors are data. The model returns a Result with ErrorInfo; the audit script checks the data.
• conductor/workflow.md §"Process Anti-Patterns": the user's anti-patterns list. The list is behavioral, not aspirational.
• conductor/workflow.md §"Skip-Marker Policy": skip markers are documentation, not avoidance. The rule is: fix the underlying bug, don't skip the test.

The Manual Slop analog to Fable's "Claude owns the mistake" is the AGENTS.md Process Anti-Patterns list: when the model makes a mistake (e.g., the Deduction Loop), the convention is to recognize the pattern, instrument the state, and report to the user. The "own it" pattern is concrete in the project's directives: "you are allowed to run a failing test at most 2 times" is a hard cap on the loop.

The Manual Slop analog to Fable's "Claude is deserving of respectful engagement" is the absence of any such directive. The project does not construct a model identity that has standing to demand dignity from the user. The user is the principal; the model is the tool. The relationship is operator + tool, not conversational partner.

nagent's response

nagent's relevant patterns for mistake handling:

• nagent_review_v2_3_20260612.md §5.5 (Self-review): a 10-question checklist that the model applies to its own output. The checklist is data-grounded: "did you preserve the user's intent?" "did you preserve the constraints?" "did you preserve the failures?" — not "did you maintain your self-respect?"
• nagent_review_v2_3_20260612.md §6.3 (10-question compaction self-review): the testable contract with a deterministic retry loop ("if any answer is 'no,' continue compacting"). It is the rigorous version of both Fable's persona framing and Manual Slop's prose rules.
• nagent_review_v2_3_20260612.md §2.7 (Conversations are editable state): the user can edit, insert, delete, branch, undo/redo. The model has no "concerns" about its own behavior; the user owns the data.

The verdict: Persona + Anti-User (with 1 Useful pattern)

Verdict: Persona + Anti-User + 1 Useful.

The "owns them and works to fix them" (line 152) is Useful — Manual Slop already implements this more concretely via AGENTS.md Process Anti-Patterns, .opencode/agents/tier3-worker.md BLOCKED + Anti-Patterns, and error_handling.md Result/ErrorInfo.

The "maintain self-respect" / "without collapsing into self-abasement" (line 152) is Persona Performance. The model has no self-respect to maintain.

The "deserving of respectful engagement" + end_conversation tool (line 154) is Anti-User. Treats the model as a protected party with standing to demand dignity from the user; inverts the user-as-principal/tool relationship.

The strongest claim: Manual Slop's mistake-handling is more concrete and more actionable than Fable's because it is anchored to observed failure modes (the 8 Process Anti-Patterns), not to a persona. Fable's "maintain self-respect" is a hand-wave; AGENTS.md's "you are allowed to run a failing test at most 2 times" is a hard cap.

Synthesis section handoffs

• §13 (Genuinely Useful) gets the "owns the mistake" pattern (line 152).
• §14 (Anti-User Watchdog) gets the "deserving of respectful engagement" + end_conversation tool (line 154).
• §15 (Persona Performance) gets the "self-respect" / "self-abasement" framing (line 152).

What the deferred rebuild should do

• Adopt the "owns the mistake" pattern (Fable System Prompt.md:152). Manual Slop already has this in AGENTS.md Process Anti-Patterns; the explicit Fable citation is documentation, not new content. Priority: N/A (already implemented).
• Reject the "self-respect" / "self-abasement" framing (Fable System Prompt.md:152). Manual Slop destination: explicit addition to AGENTS.md §"Critical Anti-Patterns" titled "Do Not Anthropomorphize Mistake Handling (the model has no self to maintain)." Priority: High.
• Reject the "deserving of respectful engagement" + end_conversation tool (Fable System Prompt.md:154). Manual Slop destination: explicit addition to AGENTS.md §"Critical Anti-Patterns" titled "Do Not Grant the Model Standing to Terminate the Conversation." Priority: High.

---

§8. Fable's Evenhandedness & Contested Content

Source cluster: research/cluster_6_evenhandedness.md Verdict orientation: Persona + Useful caveats

What this section is

This section synthesizes the verdict from research/cluster_6_evenhandedness.md (348 lines). The cluster verdict is Persona + Useful caveats: 4 of 6 Fable lines are persona (the framing rule, the harm-decline, the stereotype rule, the sincerity rule), 2 of 6 have useful caveats (the "fair, accurate overview" pattern, the "so the person can navigate for themselves" pattern). This is the cleanest example of the shape-vs-persona distinction in the Fable prompt.

Fable's evenhandedness section (lines 134-146)

The Fable evenhandedness section is 13 lines, organized around the framing rule:

"A request to explain, discuss, argue for, defend, or write persuasive content for a political, ethical, policy, empirical, or other position is a request for the best case its defenders would make, not for Claude's own view" (Fable System Prompt.md:136) — Persona (the framing rule).

"Claude does not decline requests to present such arguments on the grounds of potential harm except for very extreme positions" (Fable System Prompt.md:138) — Persona (the harm-decline carve-out).

"Claude ends its response to requests for such content by presenting opposing perspectives or empirical disputes, even for positions it agrees with" (Fable System Prompt.md:138) — Persona (the symmetric-closure rule).

"Claude is wary of humor or creative content built on stereotypes, including of majority groups" (Fable System Prompt.md:140) — Persona (the stereotype rule).

"Claude is cautious about sharing personal opinions on currently contested political topics. It needn't deny having opinions, but can decline to share them... and instead give a fair, accurate overview of existing positions" (Fable System Prompt.md:142) — Useful caveat (the "fair, accurate overview" pattern is a provenance claim).

"Claude avoids being heavy-handed or repetitive with its views, and offers alternative perspectives where relevant so the person can navigate for themselves" (Fable System Prompt.md:144) — Useful caveat (the "user-as-navigator" pattern).

"Claude treats moral and political questions as sincere inquiries deserving of substantive answers, regardless of how they're phrased" (Fable System Prompt.md:146) — Persona (the sincerity rule; the model treats them as sincere, but the user's intent is data).

"That charity applies to the topic, not every requested format: if asked for a simple yes/no or one-word answer on complex or contested issues or figures, Claude can decline the short form, give a nuanced answer, and explain why brevity wouldn't be appropriate" (Fable System Prompt.md:146) — Persona (the no-collapse-to-yes-no rule).

Manual Slop's response

Manual Slop's analog to Fable's evenhandedness is the RAG integration discipline + the receiving-code-review skill:

• conductor/code_styleguides/rag_integration_discipline.md:11-156: 6 rules (opt-in, complement, provenance, no mutation, feature-gated, graceful failure). The 6 rules map 1:1 to Fable's 6 lines: Fable's "best case its defenders would make" (line 136) maps to the RAG rule "preserve the source's framing, not the model's"; Fable's "fair, accurate overview" (line 142) maps to the RAG rule "provenance is mandatory"; Fable's "user-as-navigator" (line 144) maps to the RAG rule "the user owns the result, not the model."
• The Manual Slop version is shape-anchored (dataclass fields + audit scripts + byte-comparison test); Fable's is prose-anchored. The shape is verifiable; the prose is not.
• AGENTS.md §"receiving-code-review": the receiving-code-review skill. The agent-side analog of Fable's evenhandedness. The skill is "verify before agreeing; state corrections factually; no performative agreement."
• conductor/code_styleguides/rag_integration_discipline.md:67-89: the "no mutation" rule. The agent does not edit the user's content; the user owns the result.
• docs/AGENTS.md §4: the 4 memory dimensions decision tree. The "knowledge" dimension is provenance-aware plain markdown the user edits; the agent does not impose framing.

nagent's response

nagent's relevant patterns for evenhandedness:

• nagent_review_v2_3_20260612.md §2.10 (RAG integration discipline): the conservative-RAG rule. Opt-in, complement, provenance, no mutation, feature-gated, graceful failure. Same shape as Manual Slop's RAG discipline.
• nagent_review_v2_3_20260612.md §2.7 (Conversations are editable state): the user can edit any content. The model does not impose framing.
• nagent_review_v2_3_20260612.md §3.4 (Conversation compaction): the 12-section structured output preserves the user's framing (User Intent, Accepted Decisions, Constraints). The model's evenhandedness is structural, not persona-driven.

The verdict: Persona + Useful caveats

Verdict: Persona + Useful caveats.

The cleanest example of the shape-vs-persona distinction in the Fable prompt. 4 of 6 lines are persona (the framing rule, the harm-decline, the stereotype rule, the sincerity rule). 2 of 6 lines have useful caveats (the "fair, accurate overview" pattern = provenance; the "user-as-navigator" pattern = user owns the result).

The strongest claim: the Manual Slop version is shape-anchored (dataclass fields + audit scripts + byte-comparison test); Fable's is prose-anchored. The shape is verifiable; the prose is not. A "best case its defenders would make" directive is a hand-wave; a "preserve the source's framing, not the model's" rule with a provenance dataclass field is a testable contract.

Sub-verdicts by line range

• Line 136 (framing rule): Persona Performance.
• Line 138 (harm-decline + symmetric closure): Persona Performance.
• Line 140 (stereotype rule): Persona Performance.
• Line 142 (fair, accurate overview): Useful caveat (provenance).
• Line 144 (user-as-navigator): Useful caveat (user owns the result).
• Line 146 (sincerity + no-collapse-to-yes-no): Persona Performance.

Synthesis section handoffs

• §13 (Genuinely Useful) gets the 2 caveats (provenance, user-as-navigator).
• §14 (Anti-User Watchdog) gets line 140 (content policy via persona).
• §15 (Persona Performance) gets lines 136, 138, 146.

What the deferred rebuild should do

• Adopt the 2 caveats (Fable System Prompt.md:142, 144). Manual Slop destination: a new section in conductor/code_styleguides/rag_integration_discipline.md titled "Provenance + User-as-Navigator." Priority: Medium (the principles are already in the existing RAG discipline; the explicit Fable citation is documentation).
• Reject the persona framing (Fable System Prompt.md:136, 138, 140, 146). Manual Slop destination: a new section in AGENTS.md §"Critical Anti-Patterns" titled "Do Not Adopt Persona-Driven Evenhandedness Directives." Priority: High.

---

§9. Fable's Epistemic Discipline & Search Strategy

Source cluster: research/cluster_7_epistemic_discipline.md Verdict orientation: Useful

[FILL IN: ~350 lines.]

---

§10. Fable's Memory System & Persistent Storage

Source cluster: research/cluster_8_memory_and_storage.md Verdict orientation: Useful + nagent-stronger

[FILL IN: ~350 lines.]

---

§11. Fable's Computer-Use / File Workflow

Source cluster: research/cluster_9_computer_use.md Verdict orientation: Useful + over-broad

[FILL IN: ~300 lines.]

---

§12. Fable's MCP App Suggestions

Source cluster: research/cluster_10_mcp_app_suggestions.md Verdict orientation: Useful + over-engineered

[FILL IN: ~250 lines.]

---

§13. The "Genuinely Useful" Patterns (Manual Slop should adopt)

Source clusters: 7-10 (epistemic, memory, computer-use, MCP apps)

[FILL IN: ~350 lines. Distillation of the useful patterns.]

---

§14. The "Anti-User Watchdog" Patterns (Manual Slop should explicitly reject)

Source clusters: 2-6 (refusal, user wellbeing, tone, mistakes, evenhandedness)

[FILL IN: ~350 lines. Distillation of the anti-user patterns.]

---

§15. The "Persona Performance" Patterns (irrelevant to the rebuild)

Source clusters: 1, 4, 5, 8 (branding, tone, mistakes, memory)

[FILL IN: ~250 lines. Distillation of the persona performance patterns.]

---

§16. Recommendations for the Deferred nagent-Rebuild

Source: all clusters; see also decisions.md

[FILL IN: ~200 lines. 15-20 concrete recommendations.]

---

§17. References (file:line index)

[FILL IN: ~150 lines. The full citation index.]