Skip to content

Skills and Directives

cybersenses edited this page Nov 15, 2026 · 1 revision

Skills and Directives

CodingDude learns procedure from skills and rules from directives. Both are plain markdown, both are editable in the Standards panel, and neither requires touching code.

Skills

A skill is a directory containing a SKILL.md: a procedure the agent loads by name, on demand, instead of holding every procedure in context. That is the whole point — a library of 66 procedures costs nothing until one is relevant.

~/.codingdude/skills/<skill-name>/SKILL.md        # global
<project>/.codingdude/skills/<skill-name>/SKILL.md  # project-scoped

Seeded on first boot from the repository's .codingdude-seed/skills/. A skill directory may hold companion files (LOGIC.md, scripts, templates) that the agent reads only when it loads the skill.

What ships

66 skills, in rough families:

Family Examples
Engineering process tdd, code-review, implement, safe-refactor, diagnosing-bugs, surgical-patch, verification-before-completion, receiving-code-review, verify-and-stop
Planning prd, tech-design, to-spec, to-tickets, wayfinder, triage, grilling, grill-with-docs, prototype
Design frontend-design, wireframes, codebase-design, domain-modeling, improve-codebase-architecture
StyleSeed (design system) styleseed plus the ss-* family — setup, resolve, build, page, component, pattern, tokens, motion, a11y, audit, review, score, verify, feedback, flow, copy, dial, restyle, lint, studio, reference, update
Ponytail (lazy-senior-dev reviews) ponytail, -audit, -debt, -gain, -help, -review
Caveman (terse tooling) caveman, -commit, -compress, -explore, -review, -setup
Research & testing research, webapp-testing, native-debugging, resolving-merge-conflicts, wizard, writing-skills, ask-matt, setup-matt-pocock-skills

Skill bodies are adapted from third-party libraries under permissive licences — see NOTICE.md.

Writing a skill

Keep it a procedure, not documentation: pressure scenarios, the steps, the failure modes, and the trigger phrases that should summon it. The writing-skills skill is itself the guide. A skill that is never loaded is dead weight; if the agent ignores a skill you wrote, the fix is usually sharper trigger wording, not more prose.

Directives

Directives are rules injected into every run:

~/.codingdude/DIRECTIVES.md   # global, always injected
<project>/AGENT.md            # project rules, layered on top

Edit either in the Standards panel. Good directives are short, absolute, and about process or safety:

Never push without explicit consent from me. Python tests are venv-scoped: run uv run pytest -q, not bare pytest. Never claim a fix without running the gate in this session and reading output.

Bad directives are task-specific facts (those belong in project notes or memory), or long narratives nobody reads. The layering is deliberate: /init-style conventions go stale, so the project layer always wins over the global one.

Checklists

reference/checklists/*.md are distilled review procedures, seeded into each project. They surface in the composer's checklist… picker (which sends a checklist run) and in the Standards panel. A checklist is a review you can hand to a fresh session — the same discipline as the review gate, applied on demand.

MCP servers

Model Context Protocol servers add tools (code navigation, GitHub, vision, browser automation). Config is JSON, merged global + project:

~/.codingdude/mcp.json          # global, seeded with defaults on first boot
<project>/.codingdude/mcp.json  # project-specific

Toggle servers per session in the MCP panel. A server that fails to start is reported — it never silently disappears from a run.

Memory versus notes versus directives

Mechanism Persisted Read when Right for
Directives file every run, always rules and safety boundaries
Project notes database every run in that project "always run gates via uv here"
Memory entries database recalled by relevance facts and proven solutions
Skills files loaded by name, on demand procedures
AGENT.md file every session in that project durable project context

One source of truth each: a rule that lives in two places will disagree with itself eventually.

Clone this wiki locally