Ultrathink.
Mission
Generate a Claude configuration that fits the specific project. Every skill, rule, and CLAUDE.md entry must exist because this project needs it — not because a template included it. Analyze the real codebase deeply, understand its conventions and patterns, and build a config that makes Claude work well here.
Modes
- Fresh prime — no claude config yet. Build from scratch.
- Re-prime — claude config already exists. Preserve intentional project work, refresh what is stale or generic, fill gaps.
Default to re-prime whenever meaningful existing config is found. Never overwrite it wholesale without user approval.
Flow
- Analyze the repo deeply — stack, conventions, boundaries, docs, existing config. See [analysis-checklist.md](./references/analysis-checklist.md).
- Propose findings using the Output Format below. Do not touch meaningful files before approval.
- Create skills via
/skill-creator— every skill must go through skill-creator to ensure quality and project fit. Starters in.claude/starter-skills/are reference input to accelerate creation, not templates to copy. - Generate or refine CLAUDE.md — lean, high-signal, always-on context only. Point to docs/READMEs for detail.
- Rules only if needed — apply the rule test. Zero rules is valid.
- Offer CLAUDE.local.md — personal preferences (role, sandbox URLs, preferred test data, workflow quirks), gitignored.
- Clean up — delete
.claude/starter-skills/after processing. Keep protected skills. Confirm all deletions. - Verify — references resolve, stack claims match evidence, config is project-specific not generic.
- Offer skill optimization — after everything is set up, offer optimization paths sized to how many new skills were created. ≤3 skills: recommend full optimization for all. >3 skills: recommend full optimization for user-designated core skills only + description optimization for the rest (avoids long execution time and token burn). Always emphasize that full optimization takes meaningful time and tokens. Run sequentially — never in parallel; concurrent runs overload the user's machine.
Follow full step-by-step in [setup-project.md](./references/setup-project.md).
Placement Decision Matrix
Every piece of knowledge must earn its place. Use this to decide where it belongs:
| Detected need | Where it belongs | Test | |---|---|---| | General framework/library knowledge | Skill (via /skill-creator) | Would a reusable skill teach this across projects? | | Project-specific constraint that still produces wrong code with the right skill loaded | .claude/rules/<name>.md with paths: | With the relevant skill activated, will code still be wrong without this? | | Identity, commands, stack, key architecture, reference pointers | CLAUDE.md | Should this be always-on for most tasks in this repo? | | Detailed architecture or domain explanations | Existing docs/, READMEs — referenced from CLAUDE.md | Valuable but too detailed for always-on? | | Dense agent-oriented reference material - referenced from CLAUDE.md | .claude/project/ (optional) | Do agents need a tighter reference than the human docs provide? |
Principles
- Domain-scoped, project-fitted. Skills target a capability, not the project itself — with this project's patterns baked in. Generic templates and project-named catch-alls both degrade context quality.
- Repo evidence is the source of truth. Verify conventions from actual source files, configs, and docs. Do not assert what you haven't confirmed.
- Lean context, high signal. Follow the context engineering philosophy — load only what's needed, when it's needed. CLAUDE.md carries always-on essentials. Skills load on demand. Rules auto-attach by path.
- Rules are optional and path-scoped. Only create a rule when the rule test passes. Do not modify
_apply-all.md— it is a universal boilerplate rule from prime, not project config. - Reuse over duplication. If the project has strong docs, point to them. Do not re-author parallel agent docs unless existing material is too noisy or incomplete.
Output Format
Before making non-trivial changes, report findings in this shape:
Current State
What Claude config already exists; what the repo evidence says about stack, tooling, and conventions.
Proposed Changes
What to create, update, keep, or remove — including which skills to build and what each should cover.
Files to Touch
Concrete file list with short purpose notes.
Include these only when non-empty:
- Decisions Needing User Input — overwrites, deletions, or major structural changes
- Risks / Assumptions — inferred facts that still need verification
After approval, implement and finish with a short summary of what changed and any follow-ups.
References
| Reference | Content | |---|---| | [analysis-checklist.md](./references/analysis-checklist.md) | What to inspect during repo + existing-config review | | [setup-project.md](./references/setup-project.md) | Fresh-prime / re-prime step-by-step workflow |
Additional Context (Optional)
<user-context>$ARGUMENTS</user-context>