claude-vibe-skills

This repo is the dev-skills skill (skills/dev-skills/). SKILL.md is billed on every request of every session that loads the skill, and each reference file is read in full when it loads, so every sentence is a recurring cost.

Load the dev-skills skill at session start in this repo, before any edit: the skill's own repo goes through its gates.

Claude Code may refuse an edit or search under skills/dev-skills/ as self-modification. Don't route around it: say so, and continue once the user approves the file or adds a permission rule for the path.

Cost and logic pass — after every skill update

Any change under skills/dev-skills/ (a fix, a feature, applied lessons) is not done until this pass has run over the changed files and the files they point at or duplicate. Do it before the version bump is committed, and report it in one line per finding (or "pass clean").

Logic

• Each rule still makes sense against the rest: no two files disagree, no reference points at a moved or deleted section, no step assumes a rule that was removed.
• The change didn't open a gap: a trigger that no longer fires, a check in checks/enforce.py whose denial text no longer matches the docs.

Cost

• No redundancy. A rule lives in one place; other files point at it. A restatement in SKILL.md of something a reference file owns goes.
• No unneeded prose. Cut sentences that explain, justify or repeat without changing what Claude does. Keep the why only where a model would otherwise talk itself out of the rule.
• Right file, right trigger. Content needed only on one path (release track, one platform, one gate) lives in a file loaded only on that path. Anything new in SKILL.md must have to fire unprompted.
• Cheap to follow. A new step doesn't add round trips, questions or re-reads the session didn't need; it chains into existing calls.
• Cache-safe. Loading a skill or reading a file appends to the conversation and keeps the prompt cache. Changing the model, fast mode, loaded MCP tools or plugins mid-session (and effort, on older models and cloud providers) re-reads the whole conversation uncached. A step that asks the user for one of those does it at session start, or states that cost. The rules that must survive a compaction sit in SKILL.md's first 5,000 tokens, which is all Claude Code re-attaches.

For a large pass, the skill-pass agent (.claude/agents/, Sonnet, read-only) can do the reading and report findings to apply.

Then rebuild and validate: bash scripts/build-skill.sh && bash scripts/validate.sh.