Authoring adept skills

Authoring a good adept skill

A skill is a directory <root>/skills/<id>/ with one SKILL.md (YAML frontmatter + markdown body) and optional sidecars. adept renders it into every enabled harness, so write it once, well, and harness-neutral.

Canonical SKILL.md format

---
id: pr-review                 # ^[a-z0-9](?:[a-z0-9-]{0,48}[a-z0-9])?$ — matches the directory name
description: Use before opening a PR. Tests, security, performance.   # <= 280 chars
activation: agent             # always | globs | agent | manual
globs: []                     # REQUIRED when activation: globs (e.g. ["**/*.go"])
allowed-tools: [Read, Grep]   # carried into Claude Code
targets: []                   # empty = all enabled harnesses
tags: [review, quality]
---
# PR Review Checklist
- [ ] Tests added or updated
- [ ] No secrets in the diff

Validated against pkg/adeptschema/skill.schema.json on every load — adept sync fails loudly on a malformed skill, so let it catch your mistakes.

Pick activation deliberately — this is the whole game

activation	Fires when	Use for
agent	The model decides from the description	Most skills. Procedures, checklists, how-tos.
globs	A matching file is in context	Language/path-specific rules (**/*.go, **/*.tsx).
always	Every request	Rare. Project-wide invariants only — costs budget on every turn.
manual	Only when explicitly invoked	Heavy or destructive procedures you don't want auto-firing.

The description is the trigger — write it like one

For activation: agent, the description is the only thing the model sees when deciding whether to load the skill. Make it earn the load:

• Lead with when to apply, not what it is: "Use before opening a PR…" beats "A PR helper."
• Name concrete nouns the user will say (file types, commands, tools, error strings).
• One line, ≤ 280 chars. If you can't say when it applies in one line, the skill is too broad — split it.

Keep it portable and lean

• Harness-neutral body. Don't say "in Cursor, click…"; describe the behavior, not the UI.
• Budget-aware. Aggregator harnesses (Codex/Copilot) cap total bytes; a bloated skill gets dropped first. Tighten prose; move long reference material into references/ sidecars.
• Scan-safe. Installs run a safety scan. Avoid anything that reads like prompt injection, exfiltrates secrets, or runs opaque remote code — it will flag (critical hard-blocks).
• Sidecars (scripts/, references/, assets/) ride along to harnesses that support them; keep the SKILL.md body self-sufficient for those that don't.

Loop

adept skill add my-skill --edit   # scaffold + open $EDITOR
adept sync                        # render to every enabled harness
adept status && adept diff        # confirm it landed clean

Run adept skill --help (and adept skill add --help) for the current flags and subcommands — prefer it over memory; it never drifts. Edit canonical, adept sync, never hand-edit rendered files. See [[using-adept]] for the full CLI and [[adept-self-improve]] for turning a one-off lesson into a durable skill.