Capability Harvest Pattern
Harvest installable tools and non-installable patterns into raw snapshots; route only audited defaults into the live agent stack.
What this solves
Agents need a broad capability map before they build or improvise. The useful pattern is not to install every public skill or MCP. The useful pattern is to keep a searchable raw catalog, then promote a capability only when a task needs it and it passes the local trust/routing gate. Source: User request, 2026-06-22
The 2026-06-24 harvest was too narrow because it treated "capability" as mostly public registries. Kevin's stack needs the broader definition: a capability can be a skill, MCP, CLI, SDK pattern, template, design technique, browser verification loop, eval/review method, source-ingest rule, or automation cadence. Many of the highest-impact improvements arrive as artifacts in X bookmarks, design folders, docs, demos, comments, repos, and agent transcripts before they ever appear in a formal registry. Source: User request, 2026-06-25
Canonical command
npm run harvest:capabilities -- --skills-limit 1000 --mcp-limit 1000 --smithery-limit 1000 --smithery-search-pages 2
The command writes dated snapshots under raw/registries/agent-capabilities/YYYY-MM-DD/:
| File | Meaning |
|---|---|
skills-sh-best-<limit>.json |
Top skills.sh candidates by official/source signal and installs |
skills-sh-sitemap.json |
Long-tail skills.sh coverage from sitemap URLs |
mcp-official-registry-<limit>.json |
Active latest servers from the official MCP Registry |
mcp-smithery-marketplace-<limit>.json |
Smithery marketplace MCPs with connection counts |
mcp-smithery-search-seeds.json |
SaaS/infrastructure search fanout using the Smithery CLI and configured product names |
cli-service-seeds.json |
Curated first-party or widely accepted service CLIs |
capability-pattern-seeds.json |
Capability classes and non-registry source feeds that must be inspected for patterns, templates, demos, and workflow upgrades |
summary.json |
Counts and file paths for the run |
After a harvest, rank promotion candidates:
npm run promote:capabilities -- --limit 50
This audit is read-only. It compares harvested candidates against installed skills and wiki coverage, then surfaces high-value gaps for human/agent promotion. It also prints the capability class and non-registry source coverage matrix so the run is not silently limited to registry data.
Capability classes
config/capability-harvest-seeds.json owns the exhaustive class list. The current classes are:
| Class | What it captures | Normal promotion target |
|---|---|---|
agent-procedure-skill |
Reusable task procedures, prompt protocols, operational checklists, specialist workflows | skills/personal/<slug>/SKILL.md, Skill Registry, skill-family page, trigger evals |
tool-or-library |
Packages, SDKs, frameworks, design systems, components, animation/effects libraries | wiki/tools/, wiki/design/, skill only if procedure recurs |
mcp-service-interface |
Local/remote MCP servers for services, data, browser, files, apps | Tool page plus installed MCP only after auth/scope/security audit |
service-cli |
Shell-native service access for SaaS, infra, data, auth, deploy, observability | Service CLI Coverage, Security and Review Skills, service skill |
sdk-or-api-pattern |
Typed clients, webhooks, auth scopes, streaming, pagination, retries | wiki/tools/, wiki/architecture/, wiki/style/, service skill |
template-or-starter |
Starter kits, repo templates, app blueprints, blocks, reference implementations | wiki/tools/, wiki/design/, scaffold skill when repeated |
design-technique |
Visual/UI patterns, typography, motion, interaction details, taste sources | wiki/design/, wiki/style/, design skill routing |
browser-and-devtools-workflow |
Browser automation, screenshots, preview, UI verification, page-state capture | webapp-testing, browser routing docs, test recipes |
eval-review-security-pattern |
Quality gates, review heuristics, evals, threat modeling, anti-regression tactics | Review/security skills, tests, routing enforcement |
data-knowledge-workflow |
ETL, RAG, search, scraping, citation, artifact capture, wiki ingestion | wiki/workflows/, wiki/meta/, source-compile rules |
agent-runtime-substrate |
Sandboxes, durable sessions, queues, browsers, remote machines, control planes | wiki/architecture/, wiki/tools/, project routing |
automation-or-cadence |
Recurring checks, monitors, sync jobs, self-heal loops, freshness rhythms | automations/<slug>.md, Operational Heartbeat, state/log updates |
Coverage passes
Each harvest should inspect at least these source families when they changed since the previous run:
| Source family | What to mine |
|---|---|
| X bookmark source compile | tools, replies, demos, videos, articles, component sources, recurring agent ideas |
| Agent transcripts | repeated user requests, missed skills, routing failures, review patterns, automation candidates |
| GitHub discovery | templates, MCP servers, CLIs, agent runtimes, design systems, testing tools |
| Official docs and changelogs | API surfaces, CLI commands, deprecations, auth scope changes, best-practice shifts |
| Design folder / moodboard sources | UI galleries, component systems, fonts, motion/effects, interaction polish |
| Frontend public sources | component sourcing, visual systems, CSS techniques, design skill updates |
| Testing/browser docs | screenshot loops, canvas checks, accessibility checks, visual regression workflows |
| Security/review tooling | review gates, threat models, supply-chain checks, auth/webhook hardening |
| AI/model/runtime docs | model routing, streaming, tool calling, sandbox runtime, cost/reporting, eval loops |
| Data/search/knowledge docs | capture, indexing, ETL, citation, artifact rendering, retrieval, data movement |
Search pattern for agents
When a task needs a capability that may already exist:
- Search installed local capabilities first:
rg --files skills config | rg '<service|domain>', thenqmd search "<service> skill mcp cli". - Search skills.sh next for procedures:
npx skills find "<domain task>"; compare againstraw/registries/agent-capabilities/*/skills-sh-best-*.json. - Search MCPs for service/data access: official MCP Registry for canonical metadata, Smithery for connection/popularity signal.
- Search CLIs for shell-agent workflows: check
config/capability-harvest-seeds.json, first-party docs, package managers, and GitHub releases. - Search non-registry patterns: X bookmark enriched artifacts, design folder sources, agent transcripts, official docs/changelogs, and GitHub examples.
- Decide the artifact shape before promoting: skill, tool page, design page, workflow, architecture note, automation, trigger eval, or source-compile rule.
- Promote only after auditing source, permissions, auth, side effects, maintenance, freshness, and fit with Agent-Native CLIs.
For SaaS/infra CLI work that lacks a service-specific skill, use Security and Review Skills and Service CLI Coverage before inventing a raw API integration.
Promotion gate
A harvested capability becomes active only when it changes future behavior:
- Add or update a
wiki/tools/page, executable skill, Skill Registry row, or skill-family page. - Add or update a
wiki/design/,wiki/style/,wiki/workflows/,wiki/architecture/, orwiki/concepts/page when the capability is a pattern rather than an installable tool. - Route it in
wiki/meta/capability-routing-map.mdandwiki/SKILL-RESOLVER.md. - Add compact routing to harness projections such as
config/cursor/rules/tool-hierarchy.mdconly after the generic route is correct. - Add it to Active Stack (What to Actually Use) only if it is a recommended default.
- Add or update a trigger eval if the route should fire from natural language.
- Add or update an automation when the pattern is recurring and has a cadence.
- Add or update source-compile rules when the pattern is a capture artifact class.
- Refresh generated surfaces with Generated Surface Contract when promotion changes packs, registries, runtime links, capability coverage, redirects, or qmd.
Do not add every harvested capability to the resolver. The snapshot is the warehouse. The resolver is the shelf of tools agents should actually grab.
Quality heuristics
| Surface | Good signal | Risk signal |
|---|---|---|
| skills.sh | Official/reputable source, high installs, current GitHub repo, audit badges | Low installs, unknown owner, vague trigger, giant prompt dump |
| Official MCP Registry | Active latest version, scoped package/remotes, clear auth metadata | Deprecated server, secret-heavy auth, broad tool surface |
| Smithery | Verified, high connection count, remote support, clear tools | Zero-use server, long marketing description, unclear owner |
| CLI | First-party docs, non-interactive JSON, dry-run, stable auth | Unmaintained package, browser-only auth, destructive defaults |
| Template / starter | Clear license, current dependencies, deployable path, real example app | Screenshot-only demo, stale lockfile, hidden vendor lock-in |
| Design technique | Inspectable implementation, accessible fallback, reusable rationale | Pure vibes, text overlap risk, motion without reduced-motion path |
| Article / post | Concrete artifact, code/demo/repo, durable source, comments add signal | No implementation detail, broken links, unverified claims |
| Agent transcript | Repeated task failure or repeated Kevin request | One-off preference, no recurrence, no observable behavior change |
Updating the seed list
Append SaaS and infrastructure gaps to config/capability-harvest-seeds.json:
- Add product names to
smitheryQueriesfor MCP marketplace discovery. - Add first-party CLIs to
cliTargetswith install command, docs URL, and mutation cautions. - Add capability classes when a new kind of reusable asset is missing from the taxonomy.
- Add non-registry source feeds when useful patterns repeatedly arrive outside skills/MCP/CLI registries.
This is intentionally appendable. If Kevin asks for a product twice and it is missing, the seed list was incomplete.
Timeline
- 2026-07-01 | Added generated-surface closeout to capability promotion so harvested tools and patterns refresh packs, registries, runtime links, capability coverage, redirects, and qmd. Source: User request, 2026-07-01
- 2026-06-25 | Expanded capability harvest from a skills/MCP/CLI registry scrape into an exhaustive capability radar covering patterns, templates, design techniques, browser/testing loops, eval/review/security methods, data workflows, runtime substrates, automations, and non-registry source feeds. Source: User request, 2026-06-25
- 2026-06-22 | Added the
promote:capabilitiesaudit command plus Security and Review Skills / Service CLI Coverage as the first horizontal promotion from the harvest. Source: User request, 2026-06-22 - 2026-06-22 | Created the harvest pattern and weekly automation for skills.sh, MCP registries, Smithery, and CLI seed discovery. Source: User request, 2026-06-22