The builder pattern

The builder pattern (reference)

Every askit-build-<type> skill follows one shape.

There is no per-component “render engine”. What a component becomes on each agent is fixed by its type:

• Skills and MCP servers are portable files. The native manifests wire them up.
• Subagents and output styles are Claude-only.
• A command has no separate Codex file. On Codex, its backing skill is the invocable form.

So the shared thing is the builder skill pattern, and it is captured here once.

The shape (create mode)

1. Interview the component’s inputs (or read them from context).
2. Copy the type’s template from templates/ to the canonical file (commands/<name>.md, agents/<name>.md, skills/<name>/SKILL.md, …).
3. Author the Claude-native canonical file: frontmatter (per Standard sec 3.8 for the type) + body.
4. Register the component in library.json components.<type> as { name, path, version, tier, status }.
5. Wire any per-target manifest pointer by running npx agent-skills-toolkit gen-manifest . --write --target=all (skills/MCP are referenced by the native manifests; Claude auto-discovers commands/ and agents/).
6. Assess: npx agent-skills-toolkit evaluate . --json and iterate to 0 errors.

The shape (improve mode)

1. npx agent-skills-toolkit evaluate . --json and read the findings.
2. Fix each finding per its message (a missing maps-to, an over-scoped tool set, a description below the bar, a drifted manifest).
3. Re-run to confirm.

Cross-agent emission (what each type does)

• Skill / MCP: portable file, referenced by each native manifest (gen-manifest). Plugin-distributable on both agents.
• Subagent / output style: Claude-only (agent-targets: [claude]); no Codex artifact (Standard sec 3.3).
• Command: a new Claude commands/<name>.md; on Codex the backing skill IS the invocable form (Standard sec 3.2 parity) - no separate Codex file.
• Chain contract / AGENTS.md: agent-agnostic; one file, no per-target form.

There is no gen-<type>.mjs render engine - per-target wiring is gen-manifest’s job, and Claude-only types have no Codex artifact. Add a per-target renderer ONLY if a genuinely divergent, both-ingestible type ever appears (none in the v1 set).