● Communauté
longranger2
⚡ Instantané
claude-gpt-workflow
par longranger2 · longranger2/claude-gpt-workflow
Delegates code execution to Codex CLI while Claude orchestrates — adversarial plan review and iterative refinement between two AI models.
Three skills that split roles: Claude orchestrates strategy, Codex executes. codex delegates coding tasks to Codex CLI. /plan-review has Codex adversarially critique Claude's plan before implementation. /plan-execute runs the approved plan with back-and-forth review loops until quality passes. Three feedback loops — plan refinement, review/fix, batch execution.
Pourquoi l'utiliser
Fonctionnalités clés
- Role separation: Claude strategizes, Codex executes
- Adversarial plan review between two models
- Three feedback loops (A: plan, B: review/fix, C: batch)
- Shell-based integration via
ask_codex.sh - Quality gate on each iteration
Démo en direct
Aperçu en pratique
claude-gpt-workflow-skill.replay
▶ prêt
Cas d'usage
Usages concrets : claude-gpt-workflow
Quand l'utiliser : Before touching code on a migration or risky refactor, you want a second model to poke holes.
Déroulement
-
Draft a plan with Claude
Write a plan.md for migrating our auth module from JWT to session cookies.✓ Copié
→ Step-by-step plan written to plan.md
-
Trigger adversarial review
/plan-review plan.md✓ Copié
→ Codex returns critique with risks Claude missed
-
Iterate until both agree
Incorporate Codex's critique, re-submit, repeat until clean.✓ Copié
→ Plan reaches stable state
Résultat : A plan that has survived real adversarial review, not just self-grading.
Pièges
- Endless review loops — Cap at 3 rounds — diminishing returns after
Quand l'utiliser : A rename / signature change across 50+ files.
Déroulement
-
Plan and approve
Plan + review a rename of <old> to <new> across the codebase.✓ Copié
→ Approved plan.md
-
Execute
/plan-execute plan.md — Codex does the edits, Claude reviews each batch.✓ Copié
→ Files change in batches, each passes Claude's review
Résultat : A clean multi-file change with two sets of eyes on every batch.
Pièges
- Codex does something Claude would have rejected — The /plan-execute loop catches it — but only if quality criteria in the plan are explicit
Quand l'utiliser : A feature complex enough that a single generation won't get it right.
Déroulement
-
Write plan with acceptance criteria
Plan a feature with explicit acceptance tests, then /plan-execute.✓ Copié
→ Plan with testable criteria
-
Watch the loop
Let the review/fix loop run — summarize each iteration.✓ Copié
→ Iterations converge to passing criteria
Résultat : A feature that passes explicit acceptance, not just a commit that compiles.
Combinaisons
Associez-le à d'autres MCPs pour un effet X10
claude-gpt-workflow-skill +
gitCommit after each loop iteration so you can revert if the loop diverges
Before each /plan-execute iteration, create a git tag.✓ Copié
Plan and critiques live in the repo
Write plan.md and review-notes.md to docs/ and keep them in git.✓ Copié
Coût et limites
Coût d'exécution
- Quota d'API
- Codex API tokens — potentially 2–5x a single-model workflow since both models run
- Tokens par appel
- High — plan + review + execute
- Monétaire
- Two API subscriptions (Claude + OpenAI)
- Astuce
- Reserve this for risky work; don't use it for small one-file edits
Sécurité
Permissions, secrets, portée
Stockage des identifiants : OPENAI_API_KEY in env for Codex CLI
Sortie de données : Code leaves your machine to both providers — check corporate policy
- Two-model workflow doubles the surface that sees your source.
Dépannage
Erreurs courantes et correctifs
ask_codex.sh: command not found
Codex CLI not on PATH or not installed
Vérifier : which codex
Plan-execute loops forever
Acceptance criteria too vague — tighten the plan or cap iterations
Models disagree and plan never stabilizes
Force a tiebreaker round with a shorter scope
Alternatives
claude-gpt-workflow vs autres
| Alternative | Quand l'utiliser | Compromis |
|---|
| Single-model Claude Code | The task isn't risky enough to justify two-model overhead | Fewer adversarial signals |
| agent-skill-tdd | Discipline comes from tests, not from another model's critique | No cross-model review |