Action-oriented checklists for multi-agent pipeline design: Reflexion, critic isolation, Constitutional AI rubrics, Plan-and-Execute decomposition, Context Engineering. Load when designing or reviewing executor/critic/orchestrator roles, pipeline iteration rules, or critique rubrics.
Load this skill when the task concerns executor/critic/orchestrator design, pipeline iteration rules, BLOCKER/WARNING/SUGGESTION rubric calibration, or any question about Reflexion, LLM-as-Judge, Constitutional AI, or Context Engineering.
## Previous Attempts)| Pattern | When to Apply |
|---|
| ReAct | All executors: Reason → Act → Observe before answering |
| Plan-and-Execute | Orchestrator decomposing a multi-step task |
| Reflexion | Executor reading past failures on iteration 2+ |
| Self-Refine | Format for every critic finding (location + root cause + fix) |
| LLM-as-Judge | Justification for critic agent isolation design |
| CRITIC | Critic is allowed only read-only tools — no web calls, no writes |
| AutoGen | Termination rule: max_iterations: 5 → NEEDS_HUMAN |
| Self-Consistency | High-stakes verdict → run critic twice; escalate on disagreement |
| Constitutional AI | BLOCKER/WARNING/SUGGESTION severity definitions |
Executor prompt checklist (iteration 2+):
## Previous Attempts in TASK_CONTEXT.md before touching filesTASK_CONTEXT.md path passed as argument — NOT the content inlined into the promptOrchestrator checklist:
## Previous Attempts (no paraphrase — A1.6)NEEDS_HUMANRed flags:
## Previous Attempts missing or skippedTASK_CONTEXT.md is gitignored (session file) but orchestrator tries to commit it → see AGENTS.mdCritic must receive ONLY:
Critic must NOT receive:
Why: executor reasoning causes anchoring bias → critic evaluates the reasoning path, not the result.
Red flag: critic returns APPROVE on a clearly incomplete result → likely received executor reasoning, not just output.
Severity definitions:
| Severity | Meaning | Pipeline action |
|---|---|---|
| BLOCKER | Must be fixed before APPROVE. Spec contradiction, broken logic, security violation, missing required section. | REQUEST_CHANGES |
| WARNING | Must be fixed before APPROVE. Likely confusion for adopters, weak rules, missing example. | REQUEST_CHANGES |
| SUGGESTION | Optional. Style, clarity, minor improvement. | Does not block APPROVE |
Rubric calibration checklist:
ACKNOWLEDGED pattern (SUGGESTION-only):
ACKNOWLEDGED: SUGGESTION | <category> | <file/section> | <issue> | Deferred: <reason>
Critic must honour this on next review — do not re-raise as WARNING/BLOCKER.
Checklist:
analysis/audit, docs-only, agent-prompt-change)Anti-patterns:
Rule: no more context than needed.
| Context type | Include | Omit |
|---|---|---|
AGENTS.md | Every orchestrator run | — |
TASK_CONTEXT.md | Executor on iter 2+ | Iter 1 |
SKILL.md | Task matches skill trigger | All other tasks |
| Full file contents | Only if editing that file | Browsing |
| Conversation history | — | Always — especially never to critic |
Violation signals:
Every critic finding must have all three:
❌ Bad: WARNING: unclear prose in §3.2
✅ Good: WARNING §3.2 "Iteration rules": "iterate as needed" has no termination condition → specify max_iterations: 5 per AutoGen pattern