Quality-first Codex skills built through real use, testing, and iteration.
English · Türkçe · Install · How it works
LunaMaxxing is an adaptive Codex workflow designed to help Luna Max produce stronger results through deliberate framing, evidence gathering, alternative generation, stepwise execution, and explicit verification.
Important
LunaMaxxing does not claim that Luna becomes intrinsically equivalent to a stronger model. It improves the process around the model so that difficult tasks are less likely to end with a shallow first answer.
| Skill | Purpose | Status |
|---|---|---|
lunamaxxing |
Quality-first analysis, planning, implementation, and verification with adaptive depth | Experimental |
Cheap reasoning is useful only when the extra work is structured. LunaMaxxing adds guardrails that make longer runs purposeful:
- Explicit invocation: it runs only when the user asks for it.
- Current-session first: it preserves conversation context by default.
- Controlled worker dispatch: a pinned Luna Max CLI worker is started only when explicitly requested.
- Config-preserving dispatch: separate workers inherit user Codex configuration unless isolation is explicitly selected.
- Adaptive depth: a simple
0–5score chooses Light, Standard, or Deep reasoning. - Bounded iteration: correction rounds are capped to prevent endless polishing.
- Authority boundaries: analysis does not silently become implementation or an external action.
- Evidence-backed confidence: more prose cannot raise confidence; stronger evidence can.
- Task-specific modules: product, research, debugging, creative work, and visual QA load only when relevant.
Use the built-in skill installer:
Use $skill-installer to install lunamaxxing from
https://github.com/HakanBabus/LunaMaxxing/tree/main/skills/lunamaxxing
The installed skill becomes available on the next turn.
python "$env:USERPROFILE\.codex\skills\.system\skill-installer\scripts\install-skill-from-github.py" `
--repo HakanBabus/LunaMaxxing `
--path skills/lunamaxxingThe installer intentionally stops if a skill with the same name already exists.
Manual installation
Clone the repository, then copy the skill folder into your Codex skills directory:
git clone https://github.com/HakanBabus/LunaMaxxing.git
Copy-Item -Recurse `
.\LunaMaxxing\skills\lunamaxxing `
"$env:USERPROFILE\.codex\skills\lunamaxxing"Start a new Codex turn after installation.
Invoke the skill explicitly:
Use $lunamaxxing to analyze this product problem, choose the strongest direction,
implement it step by step, and verify the result.
The skill stays in the current session unless you explicitly request a separate worker:
Use $lunamaxxing in a separate pinned Luna Max CLI worker for this task.
Separate workers are intentionally one-shot and non-resumable. Each explicit request authorizes one ephemeral worker. They inherit your normal Codex configuration, including configured MCP servers and preferences; request isolated config explicitly when that is the desired constraint.
-OutputLastMessage accepts an absolute path or a path relative to -Workdir; the launcher creates a missing output directory only for a real worker run, never during -DryRun.
More example prompts
Use $lunamaxxing to diagnose this regression before changing any code.
Use $lunamaxxing to compare three product directions, select one, implement it,
and validate the user-visible result.
Use $lunamaxxing to research this decision, separate facts from inference,
and produce an implementation-ready plan.
- Respect authority — determine whether the request permits analysis, local changes, or external actions.
- Select the route — continue in the current session by default; dispatch only when explicitly requested.
- Score the task — use five simple uncertainty and risk signals to choose the depth.
- Frame success — define the outcome, acceptance criteria, constraints, preserved behavior, and open risks.
- Establish evidence — inspect the real system and test competing explanations.
- Explore and decide — generate materially different directions when the task warrants it.
- Execute step by step — make focused changes and inspect each result.
- Verify independently — test purpose, behavior, artifacts, regressions, and the most damaging plausible failure.
- Report confidence — tie Low, Medium, or High confidence to actual validation evidence.
| Score | Budget | Default correction limit | Typical use |
|---|---|---|---|
0–1 |
Light | 1 | Localized, deterministic, low-risk work |
2–3 |
Standard | 2 | Non-trivial work with uncertainty or alternatives |
4–5 |
Deep | 3 | Ambiguous, cross-cutting, high-risk, or heavily state-dependent work |
The limits are allowances, not targets. The workflow stops early when evidence already supports acceptance.
LunaMaxxing/
├─ .github/workflows/test.yml
├─ skills/
│ └─ lunamaxxing/
│ ├─ SKILL.md
│ ├─ agents/openai.yaml
│ ├─ references/
│ └─ scripts/
├─ tests/test-lunamaxxing.ps1
├─ README.md
├─ README.tr.md
├─ CONTRIBUTING.md
└─ LICENSE
- The skill is not authority for destructive actions, purchases, credential use, production changes, or scope expansion.
- Model identity is reported only when runtime metadata verifies it or the launcher pins it.
- A separate worker is never created merely to gain more thinking time.
- Separate workers are ephemeral one-shot runs and cannot be resumed after exit.
- User Codex configuration is inherited by default; isolated config is opt-in.
- Some CLI flags and model identifiers may depend on the user's Codex version and account availability.
- The project is experimental; inspect the workflow before using it on high-impact work.
The current version is ready for practical testing and public iteration. Planned additions include reproducible example tasks, Luna Max versus LunaMaxxing comparisons, evaluation tables, and visual result graphs.
Issues and focused pull requests are welcome. See CONTRIBUTING.md.
Released under the MIT License.