skills: add anti-rationalization sections to 4 core skills

Add Common Rationalizations (table format), Red Flags (observable
symptoms), and Verification (checkbox checklists) sections to:
- flow-code-plan
- flow-code-work
- flow-code-impl-review
- flow-code-epic-review

Follows the pattern established in flow-code-debug SKILL.md.

Task: fn-136-borrow-agent-skills-patterns-anti.3

Author: z23cc

skills/flow-code-epic-review/SKILL.md

@@ -188,3 +188,29 @@ Epic review passed. Next:
 2) Ship the code: push + create PR
 3) Start next epic: `/flow-code:plan <idea>`
 ```
+
+## Common Rationalizations
+
+| Excuse | Reality |
+|--------|---------|
+| "All tasks passed, epic must be fine" | Task-level success ≠ epic-level coherence. Integration gaps hide between tasks |
+| "Tests pass, ship it" | Tests verify code behavior, not user value. Goal-backward check verifies intent |
+| "We already reviewed each task" | Per-task review misses cross-task consistency and architectural drift |
+| "The user will catch any issues" | Your job is to catch issues before the user does |
+| "Good enough for now" | "Good enough" compounds into technical debt. Ship complete or scope down |
+
+## Red Flags
+
+- All acceptance criteria marked "Yes" but no test evidence provided
+- Implementation diverged from epic plan (compare git diff vs spec scope)
+- Goal-backward verification skipped or done superficially
+- Epic has open required gaps but no justification for proceeding
+- No cross-task integration verification performed
+
+## Verification
+
+- [ ] Every acceptance criterion verified with evidence (test output, screenshot, or manual check)
+- [ ] Goal-backward verification completed for each criterion
+- [ ] No open required gaps in gap registry
+- [ ] Epic spec updated if scope changed during implementation
+- [ ] Cross-task consistency verified (no conflicting changes between tasks)

skills/flow-code-impl-review/SKILL.md

@@ -155,3 +155,29 @@ If verdict is NEEDS_WORK, loop internally until SHIP:
 5. **Repeat** until `<verdict>SHIP</verdict>`
 
 **CRITICAL**: For RP, re-reviews must stay in the SAME chat so reviewer has context. Only use `--new-chat` on the FIRST review.
+
+## Common Rationalizations
+
+| Excuse | Reality |
+|--------|---------|
+| "Reviewer said fix, so it must be right" | Review feedback needs technical verification too. Reviewers can be wrong |
+| "This is just cosmetic" | 3 cosmetic issues often signal a structural problem underneath |
+| "3 iterations is overkill for this" | The circuit breaker exists for a reason. If you hit it, the plan was wrong |
+| "I already fixed the important ones" | "Important" is defined by severity, not your estimate of effort |
+| "The code works, reviewer is being pedantic" | Working code is the minimum bar, not the finish line |
+
+## Red Flags
+
+- Same issue appears across 3+ review iterations (underlying design problem)
+- Fix introduces a new Critical finding
+- SHIP verdict issued with zero test evidence
+- Review only checks code quality but ignores spec compliance
+- ≥3 Critical findings in a single review pass
+
+## Verification
+
+- [ ] All Critical findings addressed with specific fixes
+- [ ] No fix introduced new Critical or Important issues
+- [ ] Tests still pass after all fixes applied
+- [ ] Review receipt saved to .flow/reviews/
+- [ ] Acceptance criteria verified (not just code quality)

skills/flow-code-plan/SKILL.md

@@ -142,3 +142,31 @@ All plans go into `.flow/`:
 **After work completes** (if auto-executed):
 - All tasks done → Layer 3 adversarial review runs automatically (Phase 3j)
 - Then auto push + draft PR (Phase 5)
+
+## Common Rationalizations
+
+| Excuse | Reality |
+|--------|---------|
+| "This is obvious, skip scouts" | Scouts find patterns and reuse points you don't know about |
+| "Only one task, no DAG needed" | Even single tasks need dependency analysis and acceptance criteria |
+| "Requirements are clear, skip interview" | Your understanding ≠ user's intent. A 2-minute check prevents 2-hour rework |
+| "This is too small to plan" | Small plans still need acceptance criteria. A 3-line spec is fine |
+| "I'll refine the plan during implementation" | Implementation drift is the #1 cause of rework. Plan now |
+
+## Red Flags
+
+- Any task estimated >L size without being split
+- Scout returns "Cannot analyze" but planning continues without investigation
+- Task title contains "and" (likely two tasks combined)
+- No acceptance criteria on any task
+- Plan has zero file references from repo-scout
+- 7+ tasks in a single epic (possible over-splitting)
+
+## Verification
+
+- [ ] Every task has ≥1 testable acceptance criterion
+- [ ] Every task is ≤M size (L tasks split)
+- [ ] File refs from repo-scout included in task specs
+- [ ] DAG validated (`flowctl validate`)
+- [ ] Quick commands section exists in epic spec
+- [ ] No task title contains "and" without justification

skills/flow-code-work/SKILL.md

@@ -136,3 +136,30 @@ $FLOWCTL restart <task-id> --force
 - Don't leave tasks half-done
 - Never use TodoWrite for task tracking
 - Never create plan files outside `.flow/`
+
+## Common Rationalizations
+
+| Excuse | Reality |
+|--------|---------|
+| "Spec is good enough, skip re-anchor" | Specs change during plan-review. Always re-read before implementing |
+| "Guard passed, definitely no issues" | Guards may not cover your specific change. Manual verification still needed |
+| "Small change, no test needed" | Small changes are the #1 source of regressions |
+| "I'll test after all tasks are done" | Cross-task integration bugs are 10x harder to diagnose than per-task bugs |
+| "One more attempt will fix it" | 3+ failed attempts = architectural problem. Escalate, don't retry |
+
+## Red Flags
+
+- ≥2 tasks failed in the same wave (systemic issue, stop and investigate)
+- Same file needs 3+ fixes across iterations
+- Guard passes but manual test reveals broken behavior
+- Worker reports "spec unclear" but continues implementing
+- >100 lines changed without running any tests
+- Wave checkpoint skipped or results not reviewed
+
+## Verification
+
+- [ ] All tasks in epic are done or explicitly skipped
+- [ ] Guards pass on final state (`flowctl guard`)
+- [ ] Evidence files exist for each completed task
+- [ ] No unresolved spec conflicts
+- [ ] Wave checkpoints completed for each wave