한국어 사용자: 문서는 영어로 작성돼 있으며, 에이전트는 사용자가 쓰는 언어로 답한다.
A set of Claude Code / Codex settings that hands planning, building, and verification to the model that fits each job, and keeps evaluation evidence for the next improvement.
Planning runs at high, building at medium, and simple execution at each model's officially recommended level (Sonnet 5.5 medium, Luna high). Results are judged by a separate top-model subagent run at xhigh. Instead of trusting the working agent's explanation, the reviewer re-checks the requirements and the actual changed files first.
The desktop session mission-control kit lives in a separate repo: claudi.
| Role | Claude Code | Effort | Codex | Effort |
|---|---|---|---|---|
| Planning · deep judgment | Fable 5.1 · claude-fable-5-1 |
high | Astra · gpt-6-astra |
high |
| Builder | Opus 5.5 · claude-opus-5-5 |
medium | Sol · gpt-6-sol |
medium |
| Execution helper | Sonnet 5.5 · claude-sonnet-5-5 |
medium | Luna · gpt-6-luna |
high |
| Independent verification | Fable 5.1 · fresh subagent | xhigh | Astra · fresh subagent | xhigh |
| Evaluation record · improvement proposals | Opus 5.5 · ai-loop |
medium | Sol · ai_loop |
medium |
"One level higher" for verification is measured against planning's high. This table is this repo's operating choice. Effort names do not mean the same amount of thinking across model generations, so recalibrate when you change models (effort docs). Check model support in the official sources and on the account you run. Do not silently swap a model you cannot use.
A setup that drops the top-tier models and only tunes effort is in Token-efficient mode.
An optional setup that assumes current models are good enough, drops the top-tier models (Fable 5.1 · Astra), and only tunes effort. Role bodies are the same; install only rewrites the model and effort lines of the role files.
| Role | Claude Code | Effort | Codex | Effort |
|---|---|---|---|---|
| Planning · deep judgment | Opus 5.5 · claude-opus-5-5 |
high | Sol · gpt-6-sol |
high |
| Builder | Opus 5.5 · claude-opus-5-5 |
medium | Sol · gpt-6-sol |
medium |
| Execution helper | Sonnet 5.5 · claude-sonnet-5-5 |
low | Luna · gpt-6-luna |
medium |
| Independent verification | Opus 5.5 · fresh subagent | xhigh | Sol · fresh subagent | xhigh |
| Red team | Opus 5.5 · ai-red-team |
xhigh | Sol · ai_red_team |
xhigh |
| Evaluation record · improvement proposals | Opus 5.5 · ai-loop |
low | Sol · ai_loop |
low |
The rule of verifying one level above planning still holds. Install with --mode eco; it combines with the other options.
python3 scripts/install.py claude --user --mode eco
python3 scripts/install.py codex --user --mode eco
python3 scripts/install.py claude --user --package red-team --mode eco
claude --model claude-opus-5-5 --effort mediumRole files already installed in standard mode differ in content, so the install stops on a collision. Move the existing files and install again. If planning conflicts become frequent or reviews start missing defects, go back to standard. Savings have not been measured yet.
Requires Python 3.9 or later. Clone the repo and install only the platform you use.
git clone https://github.com/mingeonho1/ai-agent.git
cd ai-agent
python3 scripts/install.py claude --userDefault paths: the Skill goes to ~/.claude/skills/claude-agent-operations/ and the five roles to ~/.claude/agents/. The evaluation-loop Skill agent-evaluation-loop is installed too. If CLAUDE_CONFIG_DIR is set, the Skill, roles, and optional hook settings go under that folder. Start a new Claude Code session and make requests with /claude-agent-operations.
claude --model claude-fable-5-1 --effort highIf CLAUDE_CODE_EFFORT_LEVEL is set, it can override each agent's effort. Use these role settings in a session where that variable is unset. Fable 5.1 needs Claude Code 2.1.257 or later; Opus 5.5 and Sonnet 5.5 need the latest version (claude update). Model config
Run the following from the same repo. GPT-6 Sol and Luna need Codex CLI 0.157.0 or later.
python3 scripts/install.py codex --user
codex --model gpt-6-astra -c 'model_reasoning_effort="high"'The Skill goes to ~/.agents/skills/codex-agent-operations/ and the five roles to ${CODEX_HOME:-~/.codex}/agents/. The evaluation-loop Skill agent-evaluation-loop is installed too. In a new Codex session, make requests with $codex-agent-operations. The role files' model and effort values apply to subagent runs; copying the Skill alone does not change models. See codex/config.example.toml to set [agents] max_concurrent_threads_per_session = 3 and review_model. Profiles live in ~/.codex/<name>.config.toml and are selected with --profile <name> (format since 0.134.0).
Instead of a global install, you can register into a single project. Both platforms can be installed into the same project.
python3 scripts/install.py claude --project /absolute/path/to/project
python3 scripts/install.py codex --project /absolute/path/to/project| Platform | Skill | Subagents |
|---|---|---|
| Claude Code | .claude/skills/ |
.claude/agents/*.md |
| Codex | .agents/skills/ |
.codex/agents/*.toml |
Use --dry-run to check paths first. Reinstalling leaves identical files alone; if different content already exists, the installer stops instead of overwriting. If you previously installed skills/claude-agent-operations into Codex, use the new Codex Skill and disable or remove the old copy in that app.
| Claude Code | Codex |
|---|---|
A Markdown agent per role with model, effort, and the tools it needs |
A TOML agent per role with model and model_reasoning_effort |
ai-reviewer runs as a fresh named subagent, separate from the build conversation. ai-loop and ai-red-team preload rubrics/checklists via skills:; ai-reviewer starts in a fresh context with no preload or memory |
ai_reviewer runs in a fresh, non-inherited context. Every review opens a new thread with spawn_agent, never resume_agent |
| Per-task handoff documents, optional post-compaction reminder hook | A short Skill entry point that reads only the references it needs; no default hooks added |
On both platforms, independent work runs in parallel and only one agent writes a given file. Verification starts after files are stable. If fixes follow, a new reviewer re-verifies the affected scope. If a non-inherited context cannot be confirmed or separate verification did not run, report independent verification as incomplete. If the reviewer asks for more execution evidence, the execution helper runs the checks and the reviewer judges the result.
Add the option below only if Claude often loses the handoff after compaction. It keeps the other settings in your existing settings.json and adds a short reminder to the compact event of SessionStart. It does not call a model or collect full transcripts.
python3 scripts/install.py claude --user --with-context-hookai-loop / ai_loop records the reviewer's evaluation and proposes the next fix scope. A fresh independent reviewer assigns scores with evidence; a Python script computes totals and the stop decision. The main agent does the actual fixes and re-evaluation calls. This stop mechanism matches the "verification subagent" gate (so the agent that did the work does not grade itself) among the four loop-exit gates in Claude Code's official best-practices; /goal and the Stop hook are optional.
flowchart LR
A[Fix requirements & criteria] --> B[Build & required checks]
B --> C[Fresh reviewer: evidence & scores]
C --> D[Loop agent: record & decide]
D -->|Pass| E[Done & capture regression cases]
D -->|Evidence-backed fix, runs left| B
D -->|Missing evidence, stall, or out of runs| F[Report why incomplete]
| Criterion | Weight | What it checks |
|---|---|---|
| Intent fit | 30% | Does it meet the requested outcome and constraints? |
| Correctness | 30% | Do actual behavior and facts hold? |
| Verification evidence | 20% | Are claims backed by matching checks and run evidence? |
| Readability | 10% | Can the next user or worker understand and use it? |
| Efficiency | 10% | Did it avoid unnecessary changes, tool calls, and repetition? |
Each criterion is scored 0–4 against observable conditions. The default pass bar is total ≥ 85, every criterion ≥ 3, required checks pass, no confirmed high/medium findings, and independent verification confirmed. Without evidence, scores are not guessed. Criteria are fixed before starting; if they change, create a new version and baseline.
At most 3 runs, including the first evaluation. Iterations where the score moves but the same problem stays unsolved do not count as improvement. Previous scores and the builder's explanation are not passed to the new reviewer; the original requirements, criteria, current output, and verification evidence are provided every time. Per-task records go in runs/<task-id>/. Verified lessons go in the project's evals/lessons/ with applicability, reproduction steps, and expected results, and the next planning pass looks up related cases and adds them as checks in the new baseline.
These weights and limits are this repo's defaults. Until calibrated against human-judged examples, scores are provisional. The loop Skill, criteria, and scripts define the exact formats.
red-team/ installs separately. It reviews plan assumptions and code failure conditions from the opposing side and returns high/medium/low impact, evidence, confidence, and how to re-verify. It does not rewrite the plan on its own or turn suspicions into confirmed defects.
python3 scripts/install.py claude --user --package red-team
python3 scripts/install.py codex --user --package red-teamClaude uses a fresh Fable 5.1 xhigh ai-red-team; Codex uses a fresh Astra xhigh ai_red_team. Request it with /adversarial-review in Claude or $adversarial-review in Codex. Use --package all to install it together with the operations setup. --project and --dry-run are supported too.
The red team finds counterexamples; the final reviewer judges that evidence. The scope borrowed from Petri and Bloom is documented. This repo's Skill does not install either project's full framework.
web-design/ installs separately. design-review scores a page or app screen with 상/중/하 (high/medium/low) findings, a 0–100 number, and pass/fail for the poc, beta or prod bar; seo-geo-review gives a deterministic SEO + GEO score. Both are read-only, and their reports are in Korean.
python3 scripts/install.py claude --user --package web-design
python3 scripts/install.py codex --user --package web-designRequest them with /design-review and /seo-geo-review in Claude, $design-review and $seo-geo-review in Codex. Two fresh reviewers (design-reviewer / design_reviewer) run independently, a 상 counts only when both flag it, and a Python script computes the score. --package all, --project and --dry-run work as above. Scoring, overfitting safeguards, and a v1 → v7 experiment (33 → 82) are in the web-design README.
dual-account/ splits two Claude accounts by role. The main account (app/CLI) only plans, judges, and synthesizes results; token-heavy building, testing, and research run headless on the second account's CLI, which returns only the final report. Only login is separated; sessions, settings, roles, and Skills are shared with ~/.claude.
bash dual-account/setup.sh
claude2 # first time only: /login with the second account
alt-agent ai-builder <<'EOF'
Goal, target files, done criteria, report format
EOFCombined with token-efficient mode, the main account plans with Opus 5.5 medium while execution roles run on the second account. Headless runs skip permission prompts, so use them only in repos you trust. Details and caveats are in dual-account/README.md.
- fable-mode/ — Fable orchestration mode. The main agent only plans, routes, and judges; a PreToolUse gate caps direct code edits and pushes execution to subagents. Toggle with
fable on/fable off. It costs more tokens than the default setup. - extras/ — general-purpose subagents:
diff-code-reviewer(reviews only the diff) andskill-evaluator(grades a SKILL.md). - examples/patsol/ — a real company profile (skills, routines, automations) kept as a worked example of how one team customized Claude Code. Company values are filled in by answering questions; nothing company-specific is hardcoded.
- Claude Skill · Claude role settings · Optional hook
- Codex Skill · Codex role settings
- Red-team bundle · Web design review bundle
- Techniques taken from official docs and their scope
- Evaluation loop rationale and scope
- Economic news card maker is a separate repo.
Token savings and quality gains have not been measured yet. Model, effort, run time, and verification results are recorded per task to tune the setup.