Skip to content

Latest commit

 

History

History
51 lines (26 loc) · 1.56 KB

File metadata and controls

51 lines (26 loc) · 1.56 KB

Anti-Patterns

Design mistakes that cause goals to fail silently or burn budget.

Self-grading

Symptom: Agent calls update_goal(completed: true) without running tests or verifier.

Do instead: goal-verifier skill (or reviewer sub-agent). Verifier default stance: REJECT.

Vague done conditions

Symptom: /goal make auth better or done checklist with subjective items ("code is clean").

Do instead: Objective gates: npm test exits 0, rg 'legacy/auth' returns empty, coverage ≥ 80%.

Mega-goals

Symptom: One /goal spans refactor + feature + migration; endless turns.

Do instead: Run goal-scoper first. Split into sequential goals with separate GOAL.md phases.

No external state

Symptom: Objective only in chat; lost after compaction.

Do instead: Mirror /goal in GOAL.md. Agent reads it every turn.

Chat progress spam

Symptom: Long status paragraphs every turn.

Do instead: Milestones → update_goal(message: "...") only. User-facing summary at end.

Loop + goal collision

Symptom: Daily loop and active goal both edit the same files.

Do instead: Loop triages → writes STATE.md → hand off one item to /goal. Pause loop during goal if needed.

Premature blocking

Symptom: blocked_reason on first flaky test or missing credential.

Do instead: Instruct: try 3+ approaches, check deny list, then block with evidence.

Skipping budget

Symptom: 80+ turns on a 15-turn objective.

Do instead: goal-budget.md turn cap, /goal pause, log in goal-run-log.md.