skills

Framework skills

Framework knowledge artifacts: plain, self-contained markdown any agent can read. Skills carry procedure and discipline; the agent-app validate gate guarantees quality even if a skill is ignored. Loaded per run, only for the run that needs them; never bound to a session or project.

Contract

• Layout: skills/<name>/SKILL.md (+ optional reference files); skills/index.json lists every skill (name, activity, description, path, optional stack/toolkit tags) so selection is deterministic — no host, no injection.
• Frontmatter: name (matches the directory), activity (one of: creator, modify, importer, operator, walk-verify, connect), description (one line: when to load), optional stack/toolkit tags. The frontmatter is agentskills.io-compatible, so harnesses that load the common SKILL.md format (OpenClaw, Hermes, dsh, Claude Code, and others) consume these unmodified.
• Self-contained: a skill states its rules directly and points only to files an operating agent actually has — the app's own AGENT_APP.md, reference/requirements.md, and (for stack specifics) reference/blueprint.md, which the blueprint ships into the app at scaffold. A skill never cites an external framework spec it cannot open.

Getting them

These ship inside the agent-app-framework package (one package, two binaries: agent-app to build/evolve, a2app to operate), so npm i -g agent-app-framework installs the skills too:

agent-app skills                      # list them
agent-app skills --path               # where they are
agent-app skills --install .claude/skills   # into a harness that loads SKILL.md

Claude Code users can instead install the plugin (.claude-plugin/marketplace.json at the repo root), which carries these six directly.

The six activities

Skill	Activity
creator/	build a new Agent App
modify/	evolve an existing one (safe-evolve: dev → gate → walk-verify → promote)
importer/	install / import / adopt / convert an existing app
operator/	read state, operate, diagnose — no code changes
walk-verify/	independently verify a running app against its requirements (verdicts: pass / defects / incomplete / blocked; never the builder)
connect/	connect to a published Agent App (reserved until deployed mode)

Stack specifics

These six are stack-agnostic: they carry the method and discipline. Stack-specific execution knowledge (the file map of what's yours vs. system-owned, the schema/migration API, how operations are implemented, the UI kit, external calls, and the gate) lives in each blueprint's reference/blueprint.md. The blueprint ships it in its template and scaffold copies it into the app, so an operating agent always has it on disk beside reference/requirements.md. These skills direct the agent to read it wherever stack specifics apply — so the contract never has to be reverse-engineered from source.