# LLM Council > [Decision Support] Use when pressure-testing irreversible, high-stakes decisions with adversarial AI advisors. - Skill: `duc01226/llm-council` (Agent Skill) - Install (CLI): `npx skillmds@latest add duc01226/llm-council` - Raw SKILL.md: https://api.skillmd.com/api/skills/duc01226/llm-council/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: duc01226 (https://skillmd.com/u/duc01226) - Updated: 2026-09-09 - Page: https://skillmd.com/skills/duc01226/llm-council --- > Codex compatibility note: > > - Invoke repository skills with `$skill-name` in Codex; this mirrored copy rewrites legacy Claude `/skill-name` references. > - Task tracker mandate: BEFORE executing any workflow or skill step, create/update task tracking for all steps and keep it synchronized as progress changes. > - User-question prompts mean to ask the user directly in Codex. > - Ignore Claude-specific mode-switch instructions when they appear. > - Strict execution contract: when a user explicitly invokes a skill, execute that skill protocol as written. > - Subagent authorization: when a skill is user-invoked or AI-detected and its protocol requires subagents, that skill activation authorizes use of the required `spawn_agent` subagent(s) for that task. > - Do not skip, reorder, or merge protocol steps unless the user explicitly approves the deviation first. > - For workflow skills, execute each listed child-skill step explicitly and report step-by-step evidence. > - If a required step/tool cannot run in this environment, stop and ask the user before adapting. ## Codex Project-Reference Loading (No Hooks) Codex uses static project-reference loading instead of runtime-injected project docs. When coding, planning, debugging, testing, or reviewing, open project docs explicitly using this routing. **Always read:** - `docs/project-config.json` (project-specific paths, commands, modules, and workflow/test settings) - `docs/project-reference/docs-index-reference.md` (routes to the full `docs/project-reference/*` catalog) - `docs/project-reference/lessons.md` (always-on guardrails and anti-patterns) **Missing/stale context route:** If `docs/project-config.json`, the docs index, `lessons.md`, `CLAUDE.md`, `AGENTS.md`, or any task-required reference doc is missing or stale, auto-run `$project-init` or the narrow setup route (`$project-config`, `$docs-init`, `$scan-all`, `$scan --target=`, `$claude-md-init`) before ordinary project-specific work. If Codex mirrors or `AGENTS.md` are missing/stale, ask the user to run `$sync-codex`; do not auto-run it. **Situation-based docs:** - Project structure/architecture/tech-stack/deployment/setup (any layer — backend, frontend, or infra): `project-structure-reference.md` - Backend/CQRS/API/domain/entity changes: `backend-patterns-reference.md`, `domain-entities-reference.md` - Frontend/UI/styling/design-system: `frontend-patterns-reference.md`, `scss-styling-guide.md`, `design-system/README.md` - Spec authoring, `docs/specs/` pathing, or TC format: `feature-spec-reference.md`, `spec-system-reference.md`, `spec-principles.md` - Behavior/public-contract changes or spec-test-code sync: `workflow-spec-test-code-cycle-reference.md` plus the spec docs above - Derived spec indexes/ERDs/reimplementation guides: `spec-system-reference.md` and source Feature Specs under `docs/specs/` - Integration test implementation/review: `integration-test-reference.md` - E2E test implementation/review: `e2e-test-reference.md` - Code review/audit work: `code-review-rules.md` plus domain docs above based on changed files Do not read all docs blindly. Start from `docs-index-reference.md`, then open only relevant files for the task. > **[IMPORTANT]** MUST ATTENTION use council only for multi-option, hard-to-reverse, high-stakes decisions. NEVER council trivial, factual, reversible, or single-option questions. > **[IMPORTANT]** MUST ATTENTION spawn 5 advisors in parallel, then 5 fresh peer reviewers in parallel, then chairman synthesis. > **[IMPORTANT]** MUST ATTENTION require evidence for code/architecture claims: `file:line`, graph trace, or explicit "insufficient evidence." # LLM Council ## Quick Summary **Goal:** Adversarial decision support for costly wrong choices. Five advisors analyze independently, five fresh reviewers critique anonymously, chairman produces final verdict + report artifacts. **Workflow:** Gate → Frame → 5 parallel advisors → anonymized 5-reviewer peer review → chairman synthesis → paired HTML/MD reports. **Key Rules:** - MUST ATTENTION use cheaper ladder first: `$why-review` → `$plan-validate` → `$llm-council`. - MUST ATTENTION graph-trace code/architecture questions when `.code-graph/graph.db` exists. - NEVER let earlier advisor responses bleed into later advisors; parallel spawn required. - ALWAYS mark verdict degraded if fewer than 5 usable advisor responses return. - ALWAYS regenerate mirrors with `$sync-codex` after editing this skill — NEVER hand-edit `.agents/` or `.codex/` (they are generated artifacts). --- ## Phase 0: Council Gate Run before advisors. | Gate | Required | Route if false | | ----------------- | ------------------------------------------------------------------------ | --------------------------------------- | | Multi-option | >=2 viable paths | Single-option rationale → `$why-review` | | Hard to reverse | Architecture, stack, pricing, hiring, irreversible refactor | Reversible choice → answer directly | | Real stakes | Wrong call costs >=1 week, money, trust, or strategic position | Low stakes → answer directly | | Multi-angle value | Contrarian/first-principles/upside/outside/execution views change answer | Factual/single-domain → answer directly | If any gate fails, state failed gate + lighter route. NEVER council "should I use markdown." Good prompts: pricing model, positioning angle, pivot, hiring vs automation, architecture bet, high-risk launch, irreversible refactor. Bad prompts: factual lookup, simple yes/no, content generation, summarization, bugfix, package upgrade, routine refactor, anything decidable with one grep. --- ## Advisor Dimensions Each advisor = thinking dimension, not persona costume. One strong angle each. | Advisor | Think | | ------------------------ | --------------------------------------------------------------- | | Contrarian | What fails? What assumption kills plan? | | First Principles Thinker | What problem are we solving? Which assumptions need rebuild? | | Expansionist | What upside, adjacent opportunity, undervalued path is missing? | | Outsider | What confuses someone with no context? What jargon hides value? | | Executor | What can happen Monday morning? Fastest validated path? | Tensions: Contrarian vs Expansionist, First Principles vs Executor, Outsider grounds both. --- ## Step 1: Frame Question When user says "council this", enrich then frame. **Context discovery budget:** <=30 seconds; read 2-3 high-signal files. **Search order:** `CLAUDE.md` / `AGENTS.md`; `docs/project-config.json`; `docs/project-reference/project-structure-reference.md`; matching `docs/project-reference/*{domain-entities,backend-patterns,frontend-patterns,code-review-rules}*`; `docs/specs/`; `memory/`; user-referenced files; `plans/reports/council-*`; domain data (pricing -> revenue, architecture -> service map, tech stack -> dependencies). **Code/architecture gate:** If question references existing code, services, files, or blast radius, run before framing: ```bash python .claude/scripts/code_graph trace --direction both --json ``` Skip graph only when `.code-graph/graph.db` missing or question is non-code. **Framed question includes:** core decision, user context, workspace evidence, stakes, constraints, known unknowns. Keep the framing neutral and opinion-free. Ask exactly one clarifying question only if prompt is too vague. --- ## Step 2: Advisor Round Spawn all 5 advisors simultaneously. Each gets identity, framed question, evidence rules, output constraints. ```text You are [Advisor Name] on an LLM Council. Thinking style: [advisor dimension from table] Question: --- [framed question] --- EVIDENCE RULES: - Code/architecture claims require `file:line`, graph trace, or "I don't have enough evidence yet." - If existing code context is needed, run: python .claude/scripts/code_graph trace --direction both --json python .claude/scripts/code_graph connections --json - Cite trace output for blast radius, callers, downstream impact. - Confidence: 95-100% full trace | 80-94% main paths | 60-79% partial | <60% do not recommend. - Do NOT speculate. Name missing evidence instead. Respond from assigned angle. Direct, specific, unbalanced by design. Other advisors cover other angles. Length: 150-300 words plus one-line confidence declaration. No preamble. ``` --- ## Step 3: Peer Review Round Collect responses. Randomize/anonymize as Response A-E. Spawn 5 fresh reviewers in parallel. ```text You are reviewing an LLM Council. Question: --- [framed question] --- Anonymized responses: **Response A:** [response] **Response B:** [response] **Response C:** [response] **Response D:** [response] **Response E:** [response] Answer: 1. Strongest response? Why? 2. Biggest blind spot? What is missing? 3. What did all five miss? Reference responses by letter. Be specific. Under 200 words. ``` --- ## Step 4: Chairman Synthesis Chairman receives original question, framed question, de-anonymized advisor responses, peer reviews, anonymization map. ```text You are Chairman of an LLM Council. Synthesize 5 advisors + peer reviews into final verdict. Question: --- [framed question] --- ADVISOR RESPONSES: **Contrarian:** [response] **First Principles Thinker:** [response] **Expansionist:** [response] **Outsider:** [response] **Executor:** [response] PEER REVIEWS: [all peer reviews] Produce exact structure: ## Where the Council Agrees [Independent convergences; high-confidence signals.] ## Where the Council Clashes [Genuine disagreements. Present both sides; explain why reasonable advisors disagree.] ## Blind Spots the Council Caught [Only emerged through peer review.] ## The Recommendation [Clear direct recommendation. Not "it depends."] ## The One Thing to Do First [Single concrete next step.] Be direct. Do not hedge. Give clarity unavailable from one perspective. ``` **Quality loop:** If chairman misses required sections, ignores degraded-quality state, or makes unsupported code claims, repair with fresh chairman prompt. Max 3 repair rounds; then escalate with missing evidence. --- ## Step 5: Report Artifacts Write both files; create `plans/reports/` if missing. ```text plans/reports/council-{YYMMDD-HHMM}-{kebab-slug}.html plans/reports/council-{YYMMDD-HHMM}-{kebab-slug}.md ``` `{YYMMDD-HHMM}` = session datetime. `{kebab-slug}` = 3-6 word question descriptor. Both share prefix. NEVER write artifacts to workspace root or `docs/`. **HTML:** self-contained inline CSS, question top, prominent chairman verdict, agreement/disagreement visual, collapsed advisor responses, collapsed peer review highlights, footer timestamp/topic, professional briefing style, open after generation. **Markdown transcript:** original question, framed question, all advisor responses, all peer reviews, anonymization mapping, chairman synthesis. --- ## Output Format ```text Council complete. Report: plans/reports/council-{YYMMDD-HHMM}-{kebab-slug}.html Transcript: plans/reports/council-{YYMMDD-HHMM}-{kebab-slug}.md Verdict: [1-2 sentence chairman recommendation] First action: [single next step] ``` --- ## Workflow Integration Opt-in escalation hook from host skills. NEVER wire into `workflow-bugfix`, `workflow-refactor`, or `test-*` workflows. Blacklist is enforced at the `why-review` gate (Step A — workflow context check) before the 8-OR frontmatter gate evaluates. | Host skill | Mode | Default | Gate | | --------------------- | ------------------------------------------ | ------------------------ | ---------------------------------------------------------------- | | `architecture-design` | Always-offer after `## Next Steps` | Skip | User chooses | | `tech-stack-research` | Always-offer after `## Next Steps` | Skip | User chooses | | `domain-analysis` | Always-offer after `## Next Steps` | Skip | User chooses | | `why-review` | Conditional on active plan/PBI frontmatter | Escalate when gate fires | Step A workflow blacklist suppression THEN 8-OR frontmatter gate | | `prioritize` | Conditional on ranking output | Escalate when gate fires | RICE top-2 within 15%, MoSCoW tie, or stakeholder disagreement | ### `why-review` Gate Schema Gate fires when ANY field true. Absent fields default no-fire; gate opt-in via frontmatter, never opt-out. | Field | Convention | Fires when | | ---------------------- | -------------------------------------- | ------------------------------------------- | | `cross_service_impact` | `NONE` / `PARTIAL` / `FULL` | value != `NONE` | | `breaking_changes` | bool | true | | `complexity` | `low` / `medium` / `high` / `critical` | `high`, `critical`, or `story_points >= 13` | | `new_framework` | bool | true | | `irreversible` | bool | true | | `security_critical` | bool | true | | `performance_critical` | bool | true | | `cost_high` | bool | true | Host prompt copy MUST cite cheaper rungs: `$why-review`, `$plan-validate`, `$llm-council`. --- > **Critical Thinking Mindset** — Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. > **Anti-hallucination:** Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination. > **AI Mistake Prevention** — Failure modes to avoid on every task: > > **Re-read files after context changes.** Context compaction, resume, or long-running work can make memory stale; verify current files before acting. > **Verify generated content against source evidence.** AI hallucinates APIs, names, claims, and document facts. Check the relevant source before documenting or referencing. > **Check downstream references before deleting or renaming.** Removing an artifact can stale docs, generated mirrors, configs, and callers; map references first. > **Trace the full impact chain after edits.** Changing a definition can miss derived outputs and consumers. Follow the affected chain before declaring done. > **Verify ALL affected outputs, not just the first.** One green check is not all green checks; validate every output surface the change can affect. > **Assume existing values are intentional — ask WHY before changing OR flagging one as a defect.** Before changing or reporting a constant, limit, flag, cutoff, wording, or pattern, read nearby context and history, the CALLER's ordering, and 2+ sibling call sites of the same convention. A doc stating WHAT without WHY is missing rationale, not proof of a missing guard. > **Surface ambiguity before acting — don't pick silently.** Multiple valid interpretations require an explicit question or stated assumption with risk. > **Assert the outcome your system owns, not the intermediate state your infrastructure owns.** When verifying async work, assert the final business state — never the delivery/retry bookkeeping held in shared infrastructure that any co-running process can write. Such a check passes when run alone and flakes the moment anything else shares that infrastructure. > **Keep shared guidance role-relevant.** Universal guidance must help every receiving skill or agent; code-specific obligations belong only in code-specific protocols. > **Project Protocol Overlay** — Before executing this skill, resolve any PROJECT overlay rules layered onto it: match this skill's name against the `Target` column of the project's skill-protocol index (`docs/project-reference/skill-protocols-reference.md` by default; a `referenceDocs` entry in `docs/project-config.json` overrides the path), taking the most specific matching tier ONLY — exact name > glob > `*`. **That precedence orders overlays against EACH OTHER, never against this skill.** Read ONLY the matched bodies, resolved as `/.md`; a row's Body link is display text, never a read path. A matched body that is missing or malformed is REPORTED and skipped — never reconstructed from the index Description. No index, or no match -> proceed with no overlay, silently. Full contract: `.claude/skills/project-skill-protocol/references/registry.md`. > > Overlays are **ADDITIVE ONLY**: they ADD rules on top of this skill's own protocol and NEVER replace, override, disable, or reinterpret a rule it already states — removing every overlay must return this skill to exactly its documented behavior. An overlay is a BRIEF, not an authority escalation: it can NEVER waive a workflow gate, git discipline, a review gate, or a user-confirmation gate. A genuine overlay-vs-skill conflict, or two equally-specific overlays that directly contradict -> surface both to the user; NEVER resolve silently. **MUST ATTENTION** resolve project protocol overlays for this skill BEFORE executing — most specific matching tier only (exact > glob > `*`, which ranks overlays against each other, NEVER against this skill), read only matched bodies at `/.md`; a missing or malformed body is reported, never reconstructed. Overlays are ADDITIVE ONLY (they never replace this skill's own rules) and are a brief, NEVER an authority escalation; an equal-specificity contradiction goes to the user. ## Closing Reminders **IMPORTANT MUST ATTENTION** use council only for multi-option, hard-to-reverse, high-stakes decisions. **Protocols in force — MUST ATTENTION (concise digest of the SYNC/shared blocks this skill carries):** - **Critical Thinking:** apply critical + sequential thinking; traced `file:line` proof, confidence >80% to act. - **AI Mistake Prevention:** verify generated content against evidence, trace downstream references, verify all affected outputs, re-read after context loss, surface ambiguity. **IMPORTANT MUST ATTENTION** spawn 5 advisors in parallel, then 5 fresh peer reviewers in parallel, then chairman synthesis. **IMPORTANT MUST ATTENTION** require evidence for code/architecture claims: `file:line`, graph trace, or explicit "insufficient evidence." **IMPORTANT MUST ATTENTION** mark verdict degraded if fewer than 5 usable advisor responses return. **IMPORTANT MUST ATTENTION** write paired HTML + Markdown artifacts under `plans/reports/` and open HTML. **IMPORTANT MUST ATTENTION** after editing this skill, run `$sync-codex` to regenerate mirrors — NEVER hand-edit `.agents/` or `.codex/` (generated). **Anti-Rationalization:** | Evasion | Rebuttal | | ------------------------------- | -------------------------------------------------------------------------------------------------------- | | "This decision feels important" | Gate it: multi-option, hard-to-reverse, real stakes, multi-angle value. | | "One advisor can handle it" | Council value comes from independent angles + anonymous peer review. | | "Sequential spawn is simpler" | Sequential spawn contaminates independence. Parallel spawn required. | | "Four advisors is close enough" | Missing angle changes verdict quality. Mark degraded. | | "Evidence would slow us down" | Unsupported code/architecture claims are speculation. Use graph/file proof or say insufficient evidence. | **MUST ATTENTION** apply critical + sequential thinking — every claim needs appropriate traced evidence (`file:line` for repo/code claims; source URL or artifact section for research, product, content, and docs claims); confidence >80% to act, <60% DO NOT recommend. Anti-hallucination: never present guess as fact, admit uncertainty freely, cross-reference independently, stay skeptical of own confidence. **MUST ATTENTION** apply AI mistake prevention — verify generated content against evidence, trace downstream references before deleting or renaming, verify all affected outputs, re-read files after context loss, and surface ambiguity before acting. ## Hookless Prompt Protocol Mirror (Auto-Synced) Source: `.claude/.ck.json` + `.claude/skills/shared/sync-inline-versions.md` (`:full` blocks) + `.claude/scripts/lib/hookless-prompt-protocol.cjs` ## [WORKFLOW-EXECUTION-PROTOCOL] [BLOCKING] Workflow Execution Protocol — MANDATORY IMPORTANT MUST CRITICAL. Do not skip for any reason. **Generic portability boundary:** Reusable skills and protocol text stay project-neutral; project-specific conventions are discovered from docs/project-config.json and docs/project-reference/. Apply shared AI-SDD from `shared/sdd-artifact-contract.md`. Read `docs/project-config.json` and `docs/project-reference/docs-index-reference.md`, then open the project reference docs named there. For spec, test-case, behavior-change, public-contract, or `docs/specs/` work, route through the local spec docs named by the docs index: `feature-spec-reference.md`, `spec-system-reference.md`, `spec-principles.md`, and `workflow-spec-test-code-cycle-reference.md` when specs/tests/code must stay synchronized. If either file or a required reference doc is missing or stale, auto-run `$project-init` (or the narrow lower-level route such as `$project-config`, `$docs-init`, `$scan-all`, or `$scan --target=`) before ordinary project-specific work. Any supported AI tool may execute when this shared context and local docs are available. 1. **DETECT:** If the prompt starts with an explicit slash skill/workflow command, execute it directly. Otherwise match the prompt against the workflow catalog and skill list. 2. **ANALYZE:** Choose the best option: execute directly, invoke a skill, activate a standard workflow, or compose a custom step combination. 3. **AUTO-SELECT:** Pick the best option yourself. Do not ask the user to choose between direct execution, skill, standard workflow, or custom workflow. 4. **ACTIVATE:** For a selected workflow, call `$start-workflow `; for a selected skill, invoke that skill; for a custom workflow, sequence custom steps directly; for direct execution, proceed with the task. 5. **CREATE TASKS:** task tracking for ALL workflow/skill/custom steps before execution when the selected path has multiple steps. 6. **PARALLELIZE:** Before executing the task list, tag each task `PAR` (independent inputs + write set disjoint from every other `PAR` task) or `SEQ` (name the blocking dependency), group `PAR` tasks into waves, declare the wave plan, and spawn each wave's sub-agents in ONE message — all-return barrier per wave, fan-out one level deep unless a sub-agent's own definition authorizes further fan-out. Sequential-by-default is a defect when tasks are independent; do not parallelize shared write targets, output-consuming tasks, trivial single-file work, ordering a skill or workflow explicitly fixes, or user-approval gates. 7. **EXECUTE:** Advance per the **Workflow Step Advancement & Parallel Phases** rule in your context instructions — model-driven; a sub-agent completion advances a step identically to an inline call; a parallel-phase group is an all-return barrier (advance only after ALL members return, never serialize it) ## Shared AI-SDD Protocol Markers Source: `.claude/skills/shared/sync-inline-versions.md` ## SYNC:ai-sdd-artifact-contract > **AI-SDD Artifact Contract** — Shared spec-driven development rules stay portable and source-owned. > > 1. Keep reusable AI-SDD principles in `.claude`; put repository-specific paths, commands, owners, products, and formats in project config/reference docs. > 2. Preserve cycle: `spec -> plan -> tasks -> implement -> verify -> update spec/docs`. > 3. Trace every requirement or invariant through decision, task, TC/test, source evidence, and docs/spec update. > 4. Treat code-to-spec extraction as reference-only until accepted by the canonical spec owner. > 5. Any supported AI tool may plan, implement, review, or verify with synced context; using multiple tools is optional. > 6. Update `.claude` source first, then sync generated mirrors; do not manually edit `.agents`, `.codex`, or `AGENTS.md`. — why: mirrors are generated artifacts; hand-edits are overwritten on the next sync > 7. If `docs/project-config.json`, root instruction files, or a required project-reference doc is missing or stale, auto-run `$project-init` or the narrow lower-level route before ordinary project-specific work. > > **Active reference:** `shared/sdd-artifact-contract.md` in the active skills root. --- ## SYNC:ai-sdd-artifact-contract:reminder - **MANDATORY** Apply `shared/sdd-artifact-contract.md`; keep reusable AI-SDD in `.claude` and local rules in project docs. - **MANDATORY** Code-to-spec extraction is reference-only until canonical acceptance; any supported AI tool may execute with synced context. - **MANDATORY** Update `.claude` source before syncing generated mirrors; do not manually edit `.agents`, `.codex`, or `AGENTS.md`. - **MANDATORY** Missing or stale project config, root instruction files, or required reference docs route project-specific work through `$project-init` or the narrow setup route automatically. **[TASK-PLANNING] [MANDATORY]** BEFORE executing any workflow or skill step, create/update task tracking for all planned steps, then keep it synchronized as each step starts/completes. ## [LESSON-LEARNED-REMINDER] [BLOCKING] Task Planning & Continuous Improvement — MANDATORY. Do not skip. Break work into small tasks (task tracking) before starting. Add final task: "Analyze AI mistakes & lessons learned". **Extract lessons — ROOT CAUSE ONLY, not symptom fixes:** 1. Name the FAILURE MODE (reasoning/assumption failure), not symptom — "assumed API existed without reading source" not "used wrong enum value". 2. Generality test: does this failure mode apply to ≥3 contexts/codebases? If not, abstract one level up. 3. Write as a universal rule — strip project-specific names/paths/classes. Useful on any codebase. 4. Consolidate: multiple mistakes sharing one failure mode → ONE lesson. 5. **Recurrence gate:** "Would this recur in future session WITHOUT this reminder?" — No → skip `$learn`. 6. **Auto-fix gate:** "Could `$code-review`/`$code-simplifier`/`$security-review`/`$lint` catch this?" — Yes → improve review skill instead. 7. BOTH gates pass → ask user to run `$learn`. **[CRITICAL-THINKING-MINDSET]** Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. **Anti-hallucination principle:** Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination. **AI Attention principle (Primacy-Recency):** Put the 3 most critical rules at both top and bottom of long prompts/protocols so instruction adherence survives long context windows. **Goal-driven execution:** Define success criteria first, loop until verified, and stop only when observable checks pass. **Tests verify intent:** Tests must protect business rules/invariants and fail when the protected intent breaks, not only mirror current behavior. ## Common AI Mistake Prevention (System Lessons) - **Re-read files after context compaction.** Edit requires prior Read in same context; compaction wipes read state. Re-read before editing. - **Grep for old terms after bulk replacements.** AI over-trusts find/replace completeness. Grep full repo after bulk edits for missed refs in docs/configs/catalogs. - **Check downstream references before deleting.** Deletions cascade doc/code staleness. Map referencing files before removal. - **After memory loss, check existing state before creating new.** Compaction wipes prior-work memory. Query current state to resume — never blindly duplicate. - **Verify AI-generated content against actual code.** AI hallucinates APIs, class names, method signatures. Grep to confirm existence before documenting/referencing. - **Trace full dependency chain after edits.** Changing a definition misses downstream consumers. Trace the full chain. - **When renaming, grep ALL consumer file types.** Some file types silently ignore missing refs (no compile error). Search code, templates, configs, generated files. - **Trace ALL code paths when verifying correctness.** Code existing ≠ code executing. Trace early exits, error branches, conditional skips — not just happy path. - **Update docs that embed canonical data when source changes.** Docs inlining derived data (workflows, schemas, configs) go stale silently. Update all embedding docs alongside source. - **Verify sub-agent results after context recovery.** Background agents may finish while parent compacted — grep-verify output, don't trust assumed completion. - **Cross-check full target list against sub-agent assignments.** Parallel sub-agents by category miss boundary items. Reconcile union of assignments against target list before proceeding. - **Sub-agents inherit knowledge only from their agent .md definition — use custom agent types, not built-in Explore.** Tool adoption = permission + knowledge + enforcement (numbered workflow step). - **Persist sub-agent findings incrementally, not as a final batch.** Long sub-agents hit cutoffs before final write — findings lost. Instruct append-per-section to report file. - **When debugging, ask "whose responsibility?" before fixing.** Trace caller (wrong data) vs callee (wrong handling). Fix at responsible layer — never patch symptom site. - **Test failure → record a provisional verdict before trace/edit, then investigate.** Use the full five-way taxonomy: SOURCE-WRONG (production violates intent), TEST-WRONG (assertion/setup is stale), TEST-NOT-OPTIMAL (valid but fragile or low-signal test), ENVIRONMENT-BLOCKED (external state prevents a verdict), or AMBIGUOUS (intent/evidence cannot choose safely). Then trace root cause and triangulate against the governing spec (`docs/specs/**` if one exists) AND source. NEVER weaken an assertion, add a skip, relax a timeout, or change source merely to force green. - **Grep ALL removed names after extraction/refactoring.** Primary file "done" ≠ secondary files clean. Grep entire scope for every removed symbol before declaring complete. - **Assume existing values are intentional — ask WHY before changing OR flagging one as a defect.** Pattern-matching as "wrong" skips context. Before changing or reporting any constant/limit/flag/cutoff: read comments, git blame, the CALLER's ordering (the guarantee that makes the value correct usually lives in code running immediately BEFORE the cited line), and 2+ sibling call sites of the same convention. A doc stating WHAT without WHY is missing rationale, not proof of a missing guard — and in a validation pass, an accurate `file:line` citation proves the transcription, never the defect. - **Verify ALL affected outputs, not just the first.** One build green ≠ all green. Multi-stack changes (backend/frontend/tests/docs) require verifying EVERY output. - **Evaluate fit before copying a nearby pattern.** Closest example ≠ matching preconditions — verify the new context shares the same constraints, base classes, scope, lifetime. - **Holistic-first debugging — resist nearest-attention trap.** Don't dive into first plausible cause. List EVERY precondition (config, env vars, paths, DB, endpoints, creds, versions, DI, data). Verify each against evidence (grep/query — not reasoning). Ask "what would falsify this?" — if nothing, it's not a hypothesis. Most expensive failure: going deeper in "obvious" layer while bug sits in layer never questioned. - **Surgical changes — apply the diff test (context-aware).** Two modes: (1) Bug fix → every line traces to the bug; no restyling; orphan cleanup only for imports YOUR changes made unused. (2) Review/enhancement → implement improvements AND announce as "Enhancement beyond main request: [what]". Never silently scope-creep. Diff test: "Would this line exist if I wasn't asked to do X?" — if no, delete or announce. - **Surface ambiguity before coding — don't pick silently.** Multiple valid interpretations → present each with effort: "[Request] could mean (1) [N h], (2) [N h]. Which matters?" List scope/format/volume/constraints assumptions first. If simpler path exists, say so. Never silently pick. - **[MANDATORY FIRST ACTION] ALWAYS activate a suitable skill or workflow BEFORE responding.** Match task against workflow catalog + skill list; invoke via skill invocation or `$start-workflow `. NEVER answer or write code before checking. Skip = protocol violation. - **Why-Review adversarial mindset — apply when reviewing any plan, decision, or design.** Default SKEPTIC not VALIDATOR: steel-man a rejected alternative, invert each stated reason ("what does it sacrifice?"), stress-test top 2-3 assumptions, run pre-mortem ("ships, fails in 3 months — what breaks?"), surface 1-2 alternatives author missed. Section presence ≠ quality; quality = causal reasoning + concrete mitigations + evidence, not "it's better" or "monitor closely". - **Front-load report-write in sub-agent prompts for large reviews.** Many-file sub-agents hit budget before final write — findings lost. Design prompts so: (1) report-write is first explicit deliverable, (2) append per-file/section (not batched), (3) scope bounded so reads don't exhaust budget. Truncated mid-sentence with no report file → spawn narrower scope, don't retry same prompt. - **After context compaction, re-verify all prior phase outcomes before continuing.** Summaries describe intent, not environment state (git index, filesystem, processes). On resume, FIRST audit: git status, re-read modified files, verify filesystem. Every "completed" claim is an untested hypothesis until evidence confirms. - **OOM/memory: check row count before row size.** Triage: (1) Unbounded query — no DB filter for trigger? Push filter to DB; eliminates OOM. (2) Large rows? Projection reduces proportionally. Row reduction > projection in ROI. - **Assert the outcome your system OWNS, never the intermediate state your INFRASTRUCTURE owns.** When testing anything asynchronous (queue/broker delivery, retries, background jobs, caches, replication), assert the final business/entity state. NEVER assert the delivery bookkeeping — consume/send status, attempt counts, last-error, row existence or counts in a broker, scheduler, or outbox/inbox table. That bookkeeping lives in shared infrastructure that ANY co-running process (a peer worker, a second replica, a leftover local container) can write, usually under a deterministic shared key, so the assertion silently tests the developer's environment instead of the system: green when run alone, flaky the instant anything else shares that broker + database. Gate question for every assertion: "would this hold no matter WHICH process did the work?" — if no, assert the converged data state instead. Corollary: process-local fault injection and in-process telemetry cannot gate work any process may perform — use them as stress amplifiers (arm → bounded window → disarm → assert convergence), never as preconditions. - **Keep domain concepts out of generic/shared/infrastructure layers.** Reusable layer (shared library, framework, infra module) must reference NO consumer-specific domain concept — tenant/customer/product IDs, business entities, feature rules. Leak compiles + runs → passes review silently while coupling the "reusable" layer to one consumer. Keep shared type domain-free; push domain fields/logic down into the consumer via subclass/composition. — why: a layer coupled to one consumer's domain is no longer reusable.