writing-kit-skills
Skill0 starsChanged 5 months agoMITView on GitHub
Author a new kit-compatible skill — frontmatter conventions, sharp trigger descriptions, kit.yaml profile tagging, capability gates, report formats, and how to register the skill so it actually fires. Use this to extend the kit instead of forking it.
In this repository29 other files it gives its agents.
- adopt-existing-projectSkill
- adrSkill
- branch-conflict-checkSkill
- ci-watchSkill
- compound-learningsSkill
- config-auditSkill
- cost-checkSkill
- coverage-ratchetSkill
- doc-syncSkill
- eval-harnessSkill
- flaky-triageSkill
- followup-trackingSkill
- incident-captureSkill
- kit-doctorSkill
- kit-updateSkill
- logging-initSkill
- migration-checkSkill
- new-project-bootstrapSkill
- nightly-auditSkill
- observability-checkSkill
- parallel-workSkill
- perf-budgetSkill
- pr-babysitterSkill
- pre-prSkill
- prompt-regressionSkill
- releaseSkill
- session-handoffSkill
- sync-healthSkill
- using-the-kitSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 41
Write it as *situation → action*: someone (or the dispatcher in using-the-kit) reading only this line must know exactly when to invoke it.
- Distinctive: line 70
If the skill only makes sense for some project archetypes, gate the whole skill (or the specific step) on a toggle, stated in this exact …
- Distinctive: line 138
Do not tell the agent to double-check, re-verify, or spawn a subagent to review its own output — that is built in and only burns tokens.
- Distinctive: line 143
Keep it as short as the content allows.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Writing Kit Skills
The kit stays coherent because every skill follows the same conventions: it declares when it triggers, reads the project profile instead of hardcoding commands, gates itself on capabilities, and ends with a verifiable report. A skill that breaks these conventions either never fires, fires with the wrong toolchain, or reports success it can't back up. Follow this guide when adding a skill to the kit or to a project that adopted it.
1. Layout and frontmatter
One directory per skill: .claude/skills/<kebab-name>/SKILL.md, with YAML
frontmatter:
---
name: <kebab-name> # must match the directory name
description: <one sentence of WHAT it does — then the trigger: WHEN to run it>
---The description field is the single most important line you will write. Skills are surfaced by matching user intent against it. A vague description means the skill never fires. State the trigger condition explicitly:
- ❌
description: Helps with database migrations - ✅
description: Catch dangerous DB schema migrations before they hit the primary DB or any replica — verify single head, reversibility, and no unguarded destructive ops
Write it as situation → action: someone (or the dispatcher in
using-the-kit) reading only this line must know exactly when to invoke it.
Optionally state the skill's model fit in the intro paragraph: code-writing/
execution skills suit models.code / models.code_light; planning, validation,
or review skills suit models.plan. Write the role key, not a bare model ID —
the routing table lives in using-the-kit and the keys in .claude/kit.yaml
(see docs/PROFILE.md). One line of prose, not a frontmatter field.
2. Read the profile — never hardcode a toolchain
Any command that varies by project reads from .claude/kit.yaml. Tag the step
with the dotted key, and show the default as the literal:
### 3. Tests
```bash
# kit.yaml → toolchain.test
uv run pytest tests/unit/ -m "not slow and not llm" --tb=short -q
```The executing agent runs the profile's value if set; the literal shown is only
the fallback default. An empty string in the profile means skip the step and
mark it N/A. Named values (trunk_branch, alerts.channel,
capabilities.spec.file) follow the same convention. See docs/PROFILE.md.
3. Gate on capabilities
If the skill only makes sense for some project archetypes, gate the whole skill (or the specific step) on a toggle, stated in this exact form near the top so it is impossible to miss:
**Applies when** `capabilities.<name>.enabled` is true in `.claude/kit.yaml`.
If false, skip and report N/A.Reporting N/A is a feature — it lets pre-pr and using-the-kit invoke every
skill unconditionally and stay correct on any project. If no existing
capability fits, add a new toggle to kit.yaml, the five presets/*.yaml,
and docs/PROFILE.md in the same change.
4. Structure the body
The house style, in order:
- Intro paragraph — why the skill exists: the failure it prevents, in terms of cost (money, red CI, silent breakage). Not a feature list.
## Workflow(or## Checks) — numbered steps. Each step is verifiable: a command with expected output, or a judgment with explicit criteria. State the boundary in the intro — one sentence. Read-only skills say so plainly ("It reads and checks only — it never modifies files";kit-doctoris canonical). Skills that write something narrow name the write and the non-write. The boundary must be exhaustive-accurate — an over-broad "read-only" claim on a skill that also writes a baseline file, a sentinel, or a chat alert is worse than no boundary at all.- Decision tables for outcomes —
| Outcome | Action |beats prose (seecoverage-ratchetstep 3). ## Output— a fenced report template the agent must fill with real results. Every conditional item gets an explicit N/A form. End destructive or irreversible flows with a confirmation requirement, and honesty rules like: "never check a box you didn't verify."
5. Register it — an unregistered skill is dead code
A new skill isn't done until it is discoverable and maintained:
- Add a row to the trigger index in
using-the-kit/SKILL.md— the dispatcher only fires skills whose index row matches; no row, no trigger. - Add it to the README: the
What's insidetree and the workflow catalog under the right purpose group. - If
pre-prshould invoke it, add a checklist step + an output line there. - If it adds hooks or settings entries, extend
kit-doctor's checks so a broken install is caught. - If it introduces a new
kit.yamlkey: updatekit.yaml, all fivepresets/*.yaml, anddocs/PROFILE.md. - Bump
kit_versionin the source kit'skit.yamlsokit-updateoffers the new skill to adopted projects. - Run
config-audit— a skill is agent-facing config; make sure it doesn't introduce injection-shaped instructions.
6. Style rules
- Address the executing agent in the imperative ("Run", "Read", "Stop and fix").
- Long-running multi-step skills keep a progress ledger (see
docs/PROGRESS_LEDGER.md) so an interrupted run resumes instead of restarting. - Bash snippets must be copy-runnable from the repo root; prefer
python3 -cone-liners over pseudo-code for anything that parses YAML/JSON. - State blocking vs. advisory explicitly for every check, and how
gates.strictness(seedocs/PROFILE.md) changes that, if it does. - Trigger on a stated condition, not on doubt. Write "Run this when
<condition>"; never "if there is even a chance…" orYou MUSTunless the rule is genuinely absolute (secrets, destructive ops, a blocking gate). - Skills that request findings ask for every finding with confidence and severity, and triage downstream — never instruct the finder to pre-filter.
- Skills that produce a written document carry one length-calibration sentence.
- Do not tell the agent to double-check, re-verify, or spawn a subagent to review its own output — that is built in and only burns tokens. Command-backed verification steps are workflow; keep those. Fan-out guidance names a cap and a criterion: only genuinely independent, sizeable tracks; never to verify your own work.
- Keep it as short as the content allows. A skill is loaded into context when invoked — every line costs tokens, and Sonnet 5's tokenizer uses ~30% more tokens for the same text than Sonnet 4.6, so trimming redundant prose pays off more than it used to.
Output
## New Skill: <name>
- SKILL.md: written (<n> lines), description states trigger: yes
- Profile keys read: <list / none>
- Capability gate: <capabilities.X.enabled / none — always applies>
- Boundary stated: <read-only | writes: <what> | none stated>
- Model fit: <code | planning/validation | mechanical | none stated>
- Registered: using-the-kit index ▢ README tree ▢ README catalog ▢
pre-pr step ▢/N/A kit-doctor check ▢/N/A presets+PROFILE.md ▢/N/A
- kit_version bumped: <old> → <new>Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

