Multi-Model Critique

Overview

Use this skill only for complex tasks. Route multiple models through the same 4-step loop (Plan -> Execute -> Review -> Improve), then run cross-critique and synthesis to produce a higher-quality final answer than any single-model draft.

Trigger rule

Enable this skill only when the request explicitly sets complex to true (or equivalent wording such as “this is complex/deep”).

If complex is false, skip this skill and respond with normal single-model behavior.

Inputs

Collect or confirm these inputs before execution:

• - complex: boolean flag (must be true)
• INLINECODE4: user request
• INLINECODE5: list of ACP agentId values (typically 3)
• INLINECODE7: output format, language, length, deadlines, forbidden assumptions
• INLINECODE8: optional runtime controls (timeoutSec, maxRetries, maxRounds, budgetUsd)

File map (what each file does)

• - SKILL.md (this file): orchestration policy, trigger conditions, and execution sequence.
• INLINECODE14: reusable prompts for draft, critique, revision, and final synthesis (includes scoring rubric usage).
• INLINECODE15: practical OpenClaw orchestration flow using sessions_spawn, sessions_send, and sessions_history.
• INLINECODE19: machine-parseable JSON output schema for final result and per-model scoring.
• INLINECODE20: utility to generate per-model prompt files for repeated runs.
• INLINECODE21: local helper that builds a run plan JSON (model mapping, round prompts, runtime settings).

Workflow

Step 1) Parallel draft round

Spawn one ACP session per model with the same task and constraints.

Per-model requirements:

• - Follow the exact internal sequence: INLINECODE22
• Print all four sections explicitly
• End with INLINECODE23

Use sessions_spawn with runtime:"acp" and explicit agentId.

Step 2) Cross-critique round

Share peer Draft Answer outputs with each model and require structured critique:

• - Strengths
• Weaknesses
• Missing assumptions/data
• Hallucination and confidence risks
• Concrete fix suggestions

Also require ranking of peer drafts with rationale.

Step 3) Revision round

Send critique feedback back to each original model and request revision:

• - Keep INLINECODE28
• Include INLINECODE29
• End with INLINECODE30

Step 4) Final synthesis round

Integrate revised answers into one user-facing output:

• - Best final answer
• Why the synthesis is stronger than individual drafts
• Remaining uncertainties
• Optional next actions

Scoring rubric (required in critique + synthesis)

Score each draft on a 1-5 scale:

• - accuracy: factual correctness and internal consistency
• INLINECODE32: completeness against user request and constraints
• INLINECODE33: quality of assumptions and support
• INLINECODE34: usefulness for concrete decision/action

Default weighted score:
INLINECODE35

Use this score to justify rankings and the final selected direction.

Prompting resources

• - Use references/prompt-templates.md for canonical prompts.
• Use scripts/build_round_prompts.py when you need file-based prompt generation for repeated or batched runs.
• Use scripts/run_orchestration.py to generate a deterministic run-plan artifact for reproducible execution.
• Use references/orchestration-template.md for concrete OpenClaw tool-call flow.

Required user-facing output shape

1. 1. INLINECODE40
2. INLINECODE41
3. INLINECODE42
4. INLINECODE43 (optional)

When machine consumption is needed, return JSON matching references/output-schema.md.

Do not expose private chain-of-thought. Provide concise reasoning summaries only.

Failure handling

• - One model fails: continue with remaining models and note reduced diversity.
• Two or more models fail: ask whether to retry or switch to single-model mode.
• Strong disagreement remains: present competing hypotheses and state what evidence would resolve them.

Runtime defaults (recommended)

• - timeoutSec: 180 per round per model
• INLINECODE46: 1 per failed model turn
• INLINECODE47: fixed at 4 (draft, critique, revision, synthesis)
• INLINECODE48: optional hard stop when cost-sensitive