Skill Router
Use this skill to find and load on-demand skills without installing the whole library into every agent session.
Debian/Linux variant (curated shortlist, not a mirror): ../../platforms/linux/skills/skill-router/SKILL.md.
Source of Truth
Canonical skill library:
<user-home>\OneDrive\01 SKILLS\.agents\skills
Skill file convention:
<user-home>\OneDrive\01 SKILLS\.agents\skills\<kebab-name>\SKILL.md
Installed agent roots are only exposure targets. The canonical OneDrive
skills/ folder is the durable library.
Routing Workflow
When the user's request may match an on-demand skill:
- Search canonical skill names and frontmatter descriptions.
- Pick the smallest set of relevant skills.
- Open and read each selected
SKILL.mdfully before using it. - If the selected skill links to task-specific references, read only the relevant referenced files.
- If no skill clearly fits, continue normally and mention that no exact skill matched.
Use PowerShell-compatible search from this machine:
$root = "<user-home>\OneDrive\01 SKILLS\.agents\skills"
Get-ChildItem -LiteralPath $root -Directory |
ForEach-Object {
$skill = Join-Path $_.FullName "SKILL.md"
if (Test-Path -LiteralPath $skill) {
Select-String -LiteralPath $skill -Pattern "description:|name:|# " -Context 0,2
}
}
Prefer rg for targeted text search:
rg -n "deck|presentation|pptx|slides|pdf|image|video|azure|supabase|vercel" "<user-home>\OneDrive\01 SKILLS\.agents\skills" -g "SKILL.md"
Do not read secret files while routing. Skill folders should normally contain instructions, references, scripts, assets, templates, and examples, not secrets.
Common Routes
Use these as starting points, then verify by reading the actual SKILL.md.
| User intent | Candidate skills |
|---|---|
| API compatibility, serialized responses, provider/consumer contracts | api-contracts; openapi-to-typescript for generated types |
| Understand an unfamiliar repository and trace its existing behavior | codebase-onboarding; codebase-design for choosing changes |
| Record or find an architectural decision and its rationale | architecture-decision-records; cross-harness-state for current task state |
| Bounded repeated repair, agent roles, terminal/worktree coordination | bounded-agent-loop, cli-agent-router |
| Hill climb a scoreable goal, iterate until threshold or plateau | hill-climb, bounded-agent-loop, skill-judge |
| Create a brand-new local skill | skill-create, skill-authoring, skill-creator |
| Add a skill from GitHub, link, pasted text, or tool docs | skill-add, skill-create |
| Update one skill or improve skills from coding practice | skill-update, skill-authoring |
| Clean up, prune, dedupe, or reduce skill context | skill-cleanup, skill-remove |
| Remove or retire a skill | skill-remove, skill-cleanup |
| Score, review or audit skill quality | skill-judge, skill-reviewer, skill-authoring |
| Audit fleet machine skill discovery, health, and agent roots | fleet-skill-audit, skill-judge |
| Find a skill or decide what applies | skill-router, find-skills, skill-authoring, related-skill |
| Coding, refactor, review, testing | refactor, requesting-code-review, systematic-debugging, webapp-testing |
| Verify a completed change, ready for review, completion evidence | verification-loop; check specifically for SPEC drift |
| Review a finished project with the user through user stories: map what users can actually do, find edge cases, reconcile the mental model with the code | user-stories |
| Polish a somewhat-done repo end to end: map, challenge, fix, improve the frontend, verify | polish-repo; user-stories for the mapping step alone, interrogate for the challenge step alone |
| Adversarial review: stress-test a change, prove it by running it, find blind spots | interrogate, blast-radius, figure-it-out, show-me-your-work |
| Understand why and how: rationale for a design, how something works, teach it | why, how, teach |
| Stop repeating the same mistake: diagnose a recurring error pattern and fix the cause | correct |
| Write tests by behavior first, regression test before the fix | tdd |
| Sharpen writing, strip slop, structure state machines and technical docs | unslop, sharpen, technical-writing, statechart |
| Reflect on a session, write down what worked, harvest a reusable lesson | reflect |
| Python fixtures, mocks, boundary cases, unittest or pytest | python-testing; python-sdk specifically for inference.sh |
| React component, hook or form tests | react-testing; webapp-testing for real browser flows |
| React or TypeScript engineering patterns | react-dev, react-useeffect, typescript-advanced-types, vercel-composition-patterns, vercel-react-view-transitions; vercel-react-native-skills for React Native |
| Agent-facing UI components: chat, tools, widgets | agent-ui, chat-ui, tools-ui, widgets-ui; web-artifacts-builder for multi-component HTML artifacts |
| Database schema, indexes, migrations | database-schema-designer; supabase-postgres-best-practices for Postgres specifics; project-db-autodetect to detect local settings |
| Application and API security practices | security-best-practices; mcp-security-hygiene for MCP configuration |
| Dependency updates and vulnerability triage | dependency-updater |
| Pi-hole DNS, Docker resolver health, Tailscale DNS routing | homelab-pihole-dns, docker-expert |
| VirtualBox guest VM access, guestcontrol, host-only networking, guest shell traps | virtualbox-guest-access |
| New machine setup, fresh install, replicate or move an agent environment | agent-machine-bootstrap; virtualbox-guest-access for VM access, opencode-bedrock-config for provider auth |
| Two accounts of the same agent CLI on one machine, wrong-account drift, WSL versus host identity | agent-account-isolation |
| Bug investigation and root cause | bug-diagnosis, systematic-debugging, backprop |
| Merge conflicts and git surgery | merge-conflict-resolution, commit-work, git-commit |
| Take a PR to merged: autonomous full send, merge-ready full stop, file a PR, take over someone else's PR, babysit CI and reviews, modernize a stale PR, leave review comments | full-send, full-stop, file-pr, takeover, babysit-pr, modernize, leaving-pr-comments |
| Spec writing (one clear feature) | to-spec, spec, build, check |
| Ticket breakdown / work planning | to-tickets, to-spec, postplan-upload |
| Explain visually, diagrams, flows, UML/C4, comparisons, HTML reports | visual-explainer; postplan-upload for hosted delivery |
| Prototypes / feasibility spikes | prototype, wayfinder |
| Large ambiguous initiatives | wayfinder, domain-modeling, codebase-design |
| Guided step-by-step setup | wizard, azure-prepare, wrangler |
| Session continuity across agents | cross-harness-state, session-handoff, molt |
| Transfer lessons, harvest rules or reuse workflows from another project | project-transfer; cross-harness-state for active sessions |
| Repo standardization and compliance | repo-standardization, shell |
| Compare repo with competitors / market upgrades | competitive-upgrade, competitor-teardown |
| Reduce token usage, compress output or memory files | caveman, caveman-compress, caveman-help; caveman-commit for commits, caveman-review for PR feedback, caveman-stats for usage, cavecrew for compressed subagents |
| Build or repair an MCP server or MCP config | mcp-builder; mcp-health-repair for failures, mcp-security-hygiene for secrets, postgres-mcp-onboarding for Postgres MCP |
| Web or frontend design | frontend-design, web-design-guidelines, shadcn, impeccable; landing-page-design for conversion pages, email-design for email |
| Anti-slop company/landing frontend: plain copy, one system, taste gate | design-taste-frontend; impeccable for app UI and design systems |
| Lazy senior-dev review: repo audit, change review, tech-debt ledger | ponytail, ponytail-review, ponytail-audit, ponytail-debt; ponytail-gain for benchmark scores, ponytail-help for the reference card |
| Brand, theme, palette or visual identity | brand-guidelines, theme-factory, logo-design-guide, canvas-design |
| Static visual assets: covers, icons, screenshots, social images | book-cover-design, app-store-screenshots, og-image-design, character-design-sheet, slack-gif-creator, algorithmic-art |
| Charts, dashboards and data storytelling | data-visualization; pitch-deck-visuals for investor decks |
| Web performance and Core Web Vitals | web-perf |
| Cloudflare and Workers | cloudflare, wrangler, workers-best-practices, durable-objects, agents-sdk, turnstile-spin; cloudflare-email-service for email, sandbox-sdk for sandboxed execution |
| Cloudflare Zero Trust / SASE | cloudflare-one; cloudflare-one-migrations for migrating off another vendor |
| Azure: apps, deploy, diagnose, cost | azure-prepare, azure-deploy, azure-validate, azure-diagnostics, azure-cost, azure-ai, microsoft-foundry |
| Azure: infrastructure, compute, storage, quotas, inventory | azure-enterprise-infra-planner, azure-compute, azure-storage, azure-quotas, azure-resource-lookup, azure-resource-visualizer, azure-upgrade, azure-cloud-migrate, azure-compliance |
| Azure: Kubernetes and AI gateway | azure-kubernetes, azure-kubernetes-automatic-readiness, airunway-aks-setup, azure-aigateway |
| Azure: data, messaging, telemetry | azure-kusto, azure-messaging, appinsights-instrumentation |
| Microsoft Entra identity and app registration | entra-app-registration; entra-agent-id for agent identities |
| Azure OpenAI model deployment and capacity | deploy-model, preset |
| Supabase or Postgres | supabase, supabase-postgres-best-practices, postgres-mcp-onboarding |
| Vercel | deploy-to-vercel, vercel-react-best-practices, vercel-cli-with-tokens |
| Documents, PDFs, spreadsheets, decks | docx, pdf, xlsx, pptx; pitch-deck-visuals for investor decks |
| Writing and communication | technical-blog-writing, press-release-writing, case-study-writing, newsletter-curation; internal-comms for internal updates, doc-coauthoring for structured docs, product-changelog for release notes |
| Marketing and launch content | seo-content-brief, product-hunt-launch, linkedin-content, content-repurposing; customer-persona for audience research |
| Prompt engineering and model prompting technique | prompt-engineering; video-prompting-guide for video models |
| Image generation or editing | ai-image-generation, gpt-image, flux-image, background-removal, image-upscaling, product-photography, ai-product-photography; model-specific: nano-banana, nano-banana-2, qwen-image-2, qwen-image-2-pro, p-image |
| Video generation | image-to-video, google-veo, seedance, happyhorse, p-video, remotion-render; video-prompting-guide for prompting, video-ad-specs for platform specs, storyboard-creation and explainer-video-guide for planning |
| Talking-head and avatar video | ai-avatar-video, talking-head-production, p-video-avatar |
| Speech, voice and music | text-to-speech, speech-to-text, ai-voice-cloning, ai-music-generation, dialogue-audio, ai-podcast-creation; ElevenLabs-specific: elevenlabs-tts, elevenlabs-stt, elevenlabs-dialogue, elevenlabs-dubbing, elevenlabs-music, elevenlabs-sound-effects, elevenlabs-voice-changer, elevenlabs-voice-isolator |
| Multi-step AI content or automation pipelines | ai-content-pipeline, ai-automation-workflows, ai-rag-pipeline, ai-marketing-videos |
| inference.sh platform: run apps, build apps, SDKs | infsh-cli, agent-tools, building-inferencesh-apps, python-sdk, javascript-sdk, python-executor, llm-models |
| Social content | ai-social-media-content, social-media-carousel, twitter-thread-creation, youtube-thumbnail-design; twitter-automation for posting via API |
| Web search, scraping and page extraction | web-search, web-to-markdown, agent-browser |
| Claude API, model ids, pricing, tool use | claude-api |
| Agent tooling and delegation | cli-agent-router, claude-code-cli, opencode-cli, codex, agent-browser |
| OpenCode Bedrock provider auth, region pinning, per-agent model override | opencode-bedrock-config |
Maintenance Rules
When adding, deleting, renaming, pruning, or auditing skills, also use
skill-authoring.
After skill library changes:
- Ensure each new skill lives at
skills/<kebab-name>/SKILL.md. - Update
default-profile.tomlif default-installed membership changed. - Update
INDEX.mdso installed/on-demand status stays accurate. - Update this router only when a new skill changes the common routing table or a renamed/deleted skill appears in the table.
- Install this router into the default agent roots so future agents can find on-demand skills.
Do not run:
dotagents sync
Use dotagents install only when agents.toml is known to be safe and the user
explicitly wants the manifest-installed defaults refreshed.
Default Install Strategy
Keep default-installed skills lean:
- Always install this
skill-router. - Always install
skill-create,skill-add,skill-update,skill-cleanup, andskill-remove. - Always install
visual-explainerandpostplan-upload: use consistent dark/cyan HTML whenever a visual helps, even for short explanations. Use visual overview -> details -> next steps; publish when hosting is authorized. - Keep essential coding, review, debugging, Cloudflare, and local setup skills installed when the user wants automatic discovery.
- Leave media, writing, document, presentation, Azure, Supabase, and Vercel skills on demand unless the user asks to make one default again.
Use default-profile.toml as the source of truth for default profile
membership. When a new CLI agent is installed, add its skill root to
machines.toml and apply the default profile.
Do not delete on-demand skills just because they are not installed.
Machine-specific values in this document use privacy placeholders.