$ kitbash install gh:addyosmani/agent-skills/skills/code-review-and-quality ✓ installed code-review-and-quality@0.0.0 budget 6000 tokens · standing 250 · mode skill · pinned in kitbash.lock ⚠ unmanifested (SKILL.md only) — defaults applied $ kitbash compile → .claude/skills/code-review-and-quality/SKILL.md → .cursor/rules/code-review-and-quality.mdc → AGENTS.md §code-review-and-quality ⚠ agentsmd cannot lazy-load — this skill costs ~5,044 tokens standing compiled 1 skill(s) for 3 agent target(s)
↑ real session — a third-party skills.sh skill, compiled to three agents, with its hidden token cost measured. Works today.
Kitbash is the open standard, package manager, and compiler for AI agent skills — versioned, tested, and pinned like real software.
Your team runs four different coding agents. A great skill written for one is dead weight for the other three — so everyone maintains divergent rule files that silently drift apart.
.claude/skills/*/SKILL.md.cursor/rules/*.mdc.github/copilot-instructions.mdAGENTS.md.windsurfrules.clinerulesCONVENTIONS.mdGEMINI.mdA Kitbash skill declares things no other format has a field for: an enforced token budget, tool permissions, and typed artifacts so skills compose through data instead of prompt-chaining hope.
# skill.toml [skill] name = "prereview" version = "0.1.0" [context] budget = 1500 # enforced, not vibes standing = 60 # cost when idle [permissions] tools = ["read", "grep", "bash:git diff *"] network = false [artifacts] produces = ["findings@1"] consumes = ["plan@1"]
Installs pin a content hash in kitbash.lock. Updates show you the instruction diff like a code review — because that's what it is. No skill ever changes behavior silently.
Three eval tiers: static lint, instruction audit, and behavioral runs against fixture repos through headless agents. Ranked by measurement, not stars.
An agent that can't run scripts gets an instruction-only variant — with a visible build warning, never a silent downgrade.
Selection rule: if a frontier model already does it well from a bare prompt, it's not a skill. Skills earn their place through project grounding, enforcement, or composition.
Reviews your diff against your team's actual standards — mined from what reviewers on this repo really flag, not a generic checklist.
Answers "why is this code like this?" with the commit, PR thread, and issue that made it so — then records the answer durably.
Turns a red CI run into a verdict: flake, environment, or real regression — with the decisive log line and the likely culprit.
Issue → file-level implementation plan with risks and test plan, as a typed artifact downstream skills consume.
Drives the changed behavior end-to-end in the running app. Tests passing is evidence, never the verdict.
Checkpointed migration batches — resumable across sessions and teammates, every batch independently revertable.
A personalized, task-anchored codebase tour that regenerates on demand — never a stale wiki page.
Typed artifacts are stdin/stdout for agents — skills pipe into each other. A pipeline is a declared graph with gates between steps: exit codes, not vibes.