Plan — Fix PM Skill Auto-Complete Flakiness (#483)#

Summary#

The product-manager skill intermittently asks clarifying questions and signals BLOCKED:PERSISTENT for simple, well-defined issues when running under the auto directive. The fix adds explicit directive-aware auto-completion logic to the skill prompt so that DIRECTIVE=auto defaults to completing with reasonable inferences, while DIRECTIVE=careful preserves existing question-asking behavior.

Files#

Steps#

1. In SKILL.md Step 6 (Analyze Completeness), add a directive-aware gate: when DIRECTIVE=auto, classify the issue as "well-specified" if it has a clear problem statement AND explicit acceptance criteria — skip question-asking and proceed directly to finalization.
2. In SKILL.md Step 8 (Take Action), add a directive-conditional branch: when DIRECTIVE=auto and the issue is well-specified, go straight to PRD finalization (rounds ≥ 3 path) instead of posting questions. Fill gaps with reasonable inferences and note assumptions.
3. In SKILL.md Step 8, add a fallback for DIRECTIVE=auto with genuinely ambiguous issues (empty body, no problem statement at all): still ask questions even in auto mode, but limit to 1 round max before forcing finalization.
4. In role.md, add a "Do" rule: "When DIRECTIVE=auto, auto-complete well-specified issues without asking questions — fill gaps with reasonable inferences and note assumptions."
5. In role.md, add a "Do Not" rule: "Do not ask clarifying questions for issues that have a clear problem statement and acceptance criteria when running under auto directive."

Verification#

• Review modified prompt text to confirm auto directive triggers auto-completion for issues with clear problem + acceptance criteria.
• Verify careful directive path is unchanged — question-asking behavior preserved.
• Confirm genuinely ambiguous issues (empty body) still trigger questions even under auto.

Risks#

• Prompt changes are non-deterministic — Claude may still occasionally ask questions despite explicit instructions. Mitigation: use strong directive language ("MUST", "NEVER") in the prompt.
• Over-aggressive auto-completion could produce low-quality PRDs for borderline issues. Mitigation: the "genuinely ambiguous" fallback ensures empty/nonsensical issues still get questions.