Skills
A skill is a scoped instruction-and-asset bundle that activates on specific triggers within a Claude-style client. It is markdown plus frontmatter plus tests, not compiled code. Anyone who can read a markdown file can read a skill.
After this page you will know what a skill contains, how triggers work, and what the six launch skills do.
Anatomy of a skill
Section titled “Anatomy of a skill”A skill lives at packages/skills/<name>/ and ships:
SKILL.md— frontmatter (name, description, version, neurotypes, triggers, MCP dependencies, profile dependencies) plus the body that the client activates against.tests/— at least three example invocations replayed in CI to catch regressions.- Assets (optional) — diagrams, templates, prompt fragments referenced from the body.
That is the whole contract. There is no compilation step; there is no skill runtime. The client reads SKILL.md and follows it.
How triggers work
Section titled “How triggers work”A skill’s triggers block in frontmatter declares the conditions under which the client activates it. Examples:
- A slash command like
/morning-brief. - A semantic match against a phrase like “what should I do next”.
- A signal from another MCP server (e.g. the rumination detector firing).
A trigger is a hint, not a contract. The user’s client decides whether to surface the skill. If you want predictable activation, prefer explicit slash commands.
Profile awareness
Section titled “Profile awareness”Every skill respects the profile. The skill SDK (@neurodock/skill-sdk for TS, neurodock-skill for Python) exposes a profile-aware formatter so authors do not re-derive the logic for output_format, max_chunk_size, and motion. See ADR 0004 for the rationale.
If a skill is tagged with a neurotype that the user has not self-identified, the skill does not activate. Self-ID gates activation, not access — diagnosis is never required.
The six launch skills
Section titled “The six launch skills”| Skill | What it does |
|---|---|
adhd-daily-planner | Morning brief that pulls overnight changes from the cognitive graph and surfaces 1–3 things that matter today. |
audhd-context-recovery | /resume command that reconstructs yesterday’s mental state from the cognitive graph. |
ocd-decision-finalizer | On repeat-validation requests, enforces a decision-finality response grounded in prior evidence. |
hyperfocus-formatter | ”Answer First” output structure that activates during long sessions when distress signals appear. |
visual-organizer | Generates Mermaid diagrams when textual output is overwhelming or when the user asks for visual organisation. |
asd-meeting-translator | Transcript → four-section brief (my asks, others’ asks, decisions, ambiguous items with verbatim quotes). Records decisions into the cognitive graph. |
The asd-meeting-translator skill depends on mcp-translation and shipped in v0.2.1 (v0.1.0, status: beta). It is part of the same skill bundle.
Skills do not call LLMs themselves
Section titled “Skills do not call LLMs themselves”A skill is content the model reads. It composes MCP tool calls (which the client makes) into a workflow. There is no vendor SDK inside a skill. This is the same vendor-neutrality property the substrate servers enforce.
What’s next
Section titled “What’s next”- Write a skill — the contributor walkthrough.
- Substrate — the MCP servers skills compose.
- Profiles — what skills read from the user’s preferences.