| name | write-review-execute |
| description | Generate and compare exactly three code scaffolds through a write-review-execute loop, then recommend the best solution using weighted multi-criteria scoring. Use when Codex must try alternative implementations, review likely failure points, optionally run proof-of-concept execution, and propose a defended best approach. |
Write Review Execute
Use this skill to produce decision-quality coding recommendations by trialing alternatives before proposing one.
Workflow
- Run intake.
- Capture objective, target files or modules, language or toolchain, hard constraints, soft constraints, and success signals.
- Ask only high-impact clarification questions when ambiguity can change feasibility or ranking.
- Resolve unanswered ambiguity with explicit assumptions.
- Choose execution mode before recommendation.
- Ask the user to choose one mode:
theoretical-only: no command execution, recommendation based on static reasoning.
proof-of-concept: run minimal executable checks for each scaffold.
- If the user does not respond, default to
proof-of-concept when a runnable command exists. Otherwise default to theoretical-only.
- Record the selected mode and confidence impact in the output.
- Follow execution_modes.md.
- Generate exactly three materially different scaffolds.
- Produce exactly 3 alternatives with distinct architecture or implementation strategies.
- Keep each scaffold targeted. Do not over-build production features unless explicitly requested.
- For each scaffold, state assumptions and expected strengths and weaknesses.
- Review each scaffold before execution.
- Run a focused review per scaffold for correctness, edge cases, error handling, maintainability, and operational risk.
- Identify ambiguity-sensitive parts that are likely to fail at runtime.
- Patch obvious defects before scoring.
- Execute according to the selected mode.
- In
proof-of-concept mode, run the smallest command set that can validate expected behavior for each scaffold.
- Capture command, output summary, and pass or fail signal for each scaffold.
- Use
scripts/evaluate_scaffolds.py for repeatable scoring and evidence capture when applicable.
- In
theoretical-only mode, skip commands and record explicit non-executed risk notes.
- Score with weighted multi-criteria.
- Apply scoring_rubric.md.
- Use normalized weighted scoring and include at least one sensitivity note.
- Enforce hard-gate failures for alternatives that violate critical correctness constraints.
- Recommend and define fallback.
- Select one primary recommendation.
- Define explicit switch conditions where a runner-up becomes preferred.
- Include one fast follow-up validation step that can confirm or falsify the recommendation.
Output Rules
- Follow the section order in recommendation_template.md.
- Always include all three alternatives, even when one is weak.
- Keep assumptions explicit and scoped to unresolved ambiguity.
- Do not present unexecuted claims as verified behavior.
- If execution is skipped or partial, reduce confidence and explain why.
Defaults
- Alternatives count:
3
- Scoring mode:
weighted
- Execution decision: ask user (
theoretical-only or proof-of-concept)
- No-mode-response fallback:
proof-of-concept when runnable, else theoretical-only
References