Agent Skill Bundles
An agent skill bundle is a distributed capability package: one setup skill configures repo-local facts, one router skill chooses the flow, and small model-invoked skills own reusable disciplines.
Why It Matters
Single giant prompts rot. One-off skills also rot when they lack shared setup, routing, docs, and pack integration. A bundle gives agents a stable operating system for a domain without hiding every procedure inside one overloaded file.
Matt Pocock's Skills (Skills for Real Engineers) is the reference pattern: ingest the bundle, rename user-facing routes into Kevin-owned vocabulary, run setup-kevin-engineering-flow once per repo, use kevin-engineering-flow as the router, then invoke smaller skills such as Agent Engineering Skills, Frontend and Design Skills, Frontend and Design Skills, Agent Engineering Skills, and Google Workspace Skills as the flow demands.
Bundle Anatomy
| Layer | Job | Kevin artifact |
|---|---|---|
| Bundle page | Explain the source, philosophy, install path, and local adoption matrix | wiki/tools/<bundle>.md |
| Setup skill | Capture repo-local facts other skills depend on | setup-kevin-engineering-flow, Security and Review Skills |
| Router skill | Help the user pick the right flow | kevin-engineering-flow, Frontend and Design Skills |
| Model-invoked leaves | Own reusable discipline and can be reached automatically | Frontend and Design Skills, Security and Review Skills, Frontend and Design Skills |
| Reference files | Hold branch-specific depth without bloating SKILL.md |
DEEPENING.md, CONTEXT-FORMAT.md, tests.md |
| Skill registry | Make executable skills discoverable without article sprawl | Skill Registry plus wiki/concepts/*-skills.md |
| Resolver rows | Make natural language route to the right owner | Skill Resolver, Active Stack (What to Actually Use) |
| Trigger evals | Prove routing behavior | config/trigger-evals.json |
| Packs | Export the bundle into portable agent context | agent-docs/packs/*.md |
User-Invoked Vs Model-Invoked
User-invoked skills are commands Kevin intentionally calls. They should orchestrate: setup, choose a flow, grill a plan, create issues, or start implementation.
Model-invoked skills should encode reusable behavior the agent can choose automatically: bug diagnosis, TDD, domain modeling, codebase design, service-specific safety, or review discipline.
The split matters because user-invoked skills spend Kevin's memory, while model-invoked skills spend prompt routing budget every turn. Do not make every bundle skill model-invoked. Keep routers user-facing and keep reusable leaves discoverable.
Claude Code Skill Portfolio Taxonomy
The Om Patel bookmark recovers a useful Anthropic skill-library taxonomy, but it should be treated as a portfolio audit, not version authority. Current Anthropic docs define Agent Skills as filesystem-based folders of instructions, scripts, templates, and reference material loaded when relevant; Claude Code skills add command invocation, bundled skills, subagent execution, and dynamic context. Source: X/@om_patel5, 2026-03-18; Source: Anthropic Agent Skills docs, 2026-07-04; Source: Claude Code Skills docs, 2026-07-04
Use these nine types as a coverage checklist for any serious bundle:
| Skill type | Local audit question |
|---|---|
| Library and API reference | Does the bundle teach private SDKs, APIs, CLIs, and gotchas better than a generic doc link? |
| Product verification | Can the agent prove the product works through browser, CLI, API, screenshot, or build gates? |
| Data fetching and analysis | Does the skill encode query paths, schemas, metrics, and source authority instead of making the model guess? |
| Business process and team automation | Does it compress repeated team workflows into owned commands with clear inputs and exit criteria? |
| Code scaffolding and templates | Does it generate repo-shaped code from fixed skeletons while preserving natural-language constraints? |
| Code quality and review | Does it catch design, correctness, security, and maintainability failures after generation? |
| CI/CD and deployment | Does it know how this repo ships, rolls back, and reports deployment status? |
| Runbooks | Does it turn incidents and maintenance procedures into executable, verified steps? |
| Infrastructure operations | Does it safely operate cloud, containers, databases, queues, and observability surfaces with approval boundaries? |
Kevin's current library already covers much of this through Security and Review Skills, no-sus-code-doctor, agent-browser, agent-docs, commit-and-push, cloud-provider skills, and source-compile automations. The gap to watch is not skill count. It is proof coverage: product verification, runbooks, and infra-ops skills need artifacts that show the work actually completed.
Cleanup Rule
When converting local skills into this pattern:
- Pick the bundle owner page.
- Add or update one setup skill if repo-local facts are required.
- Add a router only if users need a flow selector.
- Keep each
SKILL.mdshort and move branch-only depth into reference files. - Regenerate Skill Registry, update the relevant skill-family page, and add resolver rows.
- Regenerate packs instead of hand-editing generated docs.
- Run
npm run skills:check,npm run routing-doctor,npm run agent-packs:refresh, andnpm run build-index.
Timeline
- 2026-07-04 | Added the Anthropic/Claude Code nine-type skill portfolio taxonomy as an audit checklist for local bundles: reference, verification, data, team process, scaffolding, review, CI/CD, runbooks, and infra ops. Source: X/@om_patel5, 2026-03-18; Anthropic Agent Skills docs, 2026-07-04
- 2026-06-27 | Created after importing the full
mattpocock/skillsengineering folder and adopting its setup/router/leaf/reference-file pattern as the cleanup model for Kevin's distributed skills. Source: User request, 2026-06-27; github.com/mattpocock/skills