[titanwings]

Requested by Tianyi · project thread

Summary

Before: hosts loaded all 782 lines of SKILL.md on every trigger, so an owner's first run read about 700 lines it never used. Colleague prompts sat in the prompts/ root (so prompts/intake.md looked generic), and the two celebrity research prompts were near-copies.

After: SKILL.md is a 41-line router. On first run the owner path reads the router plus a 19-line prompts/flows/owner.md. Every prompt has a folder named after who uses it, and there is one research prompt.

Stacked on #206. It should be retargeted to dot-skill once #206 merges.

Changes

• Tests first: tests/test_skill_entrypoint_docs.py no longer pins ~35 phrases to SKILL.md. It now checks the contract: the frontmatter name and invocations, that every bundled path the instructions mention exists, that every prompt is loaded by something, that the router stays under 80 lines and reaches every flow, and that the documented research floors are metrics merge_research.py actually reports. The other tests in that file are unchanged.
• Router + flows: SKILL.md keeps the skill-root contract, a request→flow table and the rules every flow shares (storage, writes through skill_writer.py, credentials). Flows live in prompts/flows/: owner, create, sources, celebrity_research, evolve, manage.
• No restating prompts: the flows point to the prompts instead of copying them (intake questions, persona builder list, research taste rules). This removes the wrong "4-question celebrity intake" claim; the intake has 5. Import options are now in A-E order, and list/rollback/delete is one {character} template.
• Folders: prompts/colleague/ holds the colleague intake, analyzer, builder and merger. prompts/shared/ holds the work analyzer, work builder and correction handler. references/ is folded into prompts/celebrity/budget_unfriendly/ (framework.md, template.md) and dropped from package.json / bin/distilly.mjs.
• One research prompt: prompts/celebrity/research.md holds the shared principles, hierarchy, blacklist, tooling and six dimension briefs once. Where the two copies differed it uses the fuller wording. It ends with a separate output contract per budget, and each keeps its own files, floor and checkpoint.

Motivation

The project goal is a light first-run install for personal agents, and the user asked what in the repo was redundant or inelegant.

Testing

• python -m unittest discover -s tests -p 'test_*.py' passed (114 tests)
• python -m compileall tools/ passed
• Manually tested:
  • Installed with install_distilly.py --host codex and --host openclaw into a temp HOME. Every flow and owner prompt the router names exists in the installed copy, and there is no references/.
  • Ran the owner first run from the installed copy: owner_sources.py discover, then owner_profile.py save, link --dry-run and link for Claude Code and OpenClaw (dry run), then unlink.
  • Created a colleague skill with skill_writer.py. Its meta.json now records prompts/colleague/* and prompts/shared/*.
  • node bin/distilly.mjs --check-package and npm pack --dry-run include prompts/flows/.

Checklist

• Docs updated if behavior or usage changed (SKILL.md)
• No secrets, tokens, or personal data committed
• Tests added/updated for new functionality

Note: skills generated before this change keep the old prompt paths in their meta.json. Nothing reads those paths back, so they still work.

🤖 Generated with Claude Code

https://claude.ai/code/session_017tXoVWiktX2vSUJzrrdBLJ

---

Generated by Claude Code