For AI agents: the complete documentation index is available at https://scriptedalchemy.github.io/agent-bundle/llms.txt, the full documentation bundle is available at https://scriptedalchemy.github.io/agent-bundle/llms-full.txt, and this page is available as Markdown at https://scriptedalchemy.github.io/agent-bundle/examples/skills-starter.md.
agent-bundle
  • English

      • Skills Starter

      A Skills-only plugin. It proves that a bundle worth installing can be authored without declaring anything but its identity and its targets — and that Skill quality can still be backed by recorded evidence rather than a claim.

      • Run from the repository root: pnpm example:skills
      • Package: @agent-bundle-example/skills-starter
      • Public dependencies: agent-bundle (workspace:*)
      • Targets: portable, codex, claude
      • Credentials: none — both eval suites are deterministic and read checked-in fixtures
      • Source: examples/skills-starter

      What it proves

      • Convention discovery. agent-bundle.config.ts declares the plugin and its three targets and nothing else. Every src/skills/*/SKILL.md directory is found by convention, so the config never lists a Skill. That is the authoring model described in Skills.
      • One source, three host layouts. The same authored documents are lowered into the portable, Codex, and Claude trees. The Workbench's Source and Generated views show whether a target copied a document or adapted it.
      • Deterministic eval evidence. The eval suites run through the deterministic harness against checked-in fixtures, so a Skill's coverage is a recorded run rather than an assertion. Coverage is labeled indirect because a deterministic harness cannot observe host Skill activation.

      What is authored

      PathWhat it is
      src/skills/incident-triage/A production incident from first signal through containment, evidence collection, and a handoff-ready update.
      src/skills/dependency-upgrade/Dependency upgrade planning with API, runtime, rollout, and rollback checks.
      src/skills/release-review/The evidence, severity, workflow, and final-report requirements for an explicit release review.
      evals/release-readiness.eval.tsThe release-readiness suite and its deterministic release-artifact-is-ready case.
      evals/engineering-operations.eval.tsThe engineering-operations suite: incident-handoff-is-actionable and upgrade-plan-has-rollback.

      Each Skill links its own references/ checklist or runbook and a reusable assets/ handoff or planning template, so the Skill body stays short and the long-form material is loaded only when it is needed.

      Working in the Workbench

      1. Overview opens on the Bundle dashboard: the three Skills, the generated targets, build health, and the next useful actions.
      2. Skills lists dependency-upgrade, incident-triage, and release-review. Browse their linked checklists and report templates, and switch between Source and Generated per target.
      3. Artifacts defaults to the Claude target. Change the target to compare the portable, Codex, and Claude output trees and their provenance.
      4. Evals defaults to the release-readiness suite. Run release-artifact-is-ready and inspect the passing trial; it consumes only the checked-in evidence fixture.
      5. To practice repair, make a reversible edit to the release policy, press Rebuild, and wait for a Failed or Idle result rather than a Building state. Restore the checked-in policy and rebuild. The earlier eval is now stale for the changed build — rerun release-readiness to record current, repaired evidence.

      Noninteractive checks

      After the repository-level pnpm build:

      cd examples/skills-starter
pnpm validate
pnpm build
pnpm exec agent-bundle eval --case release-artifact-is-ready --trials 1

      pnpm check is the validation-and-build pair without the Workbench. Generated output lands in artifact/, and its root contract is artifact/agent-bundle.manifest.json. The .agent-bundle/ directory holds development state and is not source material.