Authoring discipline for new IJFW skills. Use when creating a new skill, writing a skill, making a skill that does X, or asked to add a skill. Trigger: create a new skill, write a skill, new skill, make me a skill, skill that does, /ijfw-writing-skills
Authoring discipline for new IJFW skills. Use when creating a new skill, writing a skill, making a skill that does X, or asked to add a skill. Trigger: create a new skill, write a skill, new skill, make me a skill, skill that does, /ijfw-writing-skills
since
1.5.0
IJFW Writing Skills -- author a new skill correctly
A skill is a thin loadable trigger surface. The description: is what Claude reads to decide whether to hot-load the body. Get the description wrong and the skill never fires; get the body wrong and the skill misfires when it does fire. This file teaches you to do both right on the first try.
Iron rule: a skill IS its trigger. If the description does not match the words a user would naturally say, the skill does not exist -- it is documentation no one reads.
1. Discovery -- the description is the trigger surface (CSO)
The description: field is the only thing Claude sees at routing time. Treat it like a search-engine snippet, not like a definition.
Shape: start with "Use when ...", end with Trigger: <verbatim phrases>, /<command-name>.
Forbidden words (enforced by scripts/lint/check-skill-descriptions.sh, case-insensitive whole-word): step, steps, phase, phases, mode, modes, workflow, stage, process, sub-skill, skill-name. These signal "the description summarises the body" and cause Claude to skip the body and follow the description shortcut.
Hard cap: 1024 characters. Aim for under 300.
# BAD -- defines the topic, uses forbidden word "process"description:"Skill for the code review process. Reviews PRs."# BAD -- workflow summary, uses "phases" and "steps"description:"Use when reviewing code -- runs three phases, fifteen steps, two passes."# GOOD -- triggers only, verbatim phrases, slash-command aliasdescription:"One-line code review comments. Trigger: review, code review, PR review, /ijfw-review"# GOOD -- "Use when X" + trigger listdescription:"Authoring discipline for new IJFW skills. Use when creating a new skill, writing a skill. Trigger: create a new skill, new skill, /ijfw-writing-skills"
2. Frontmatter -- the loadable contract
Required:
name: -- letters, digits, hyphens only. Must match the directory name. Prefix with ijfw- for IJFW skills.
description: -- see section 1.
Optional but recommended:
since: "1.5.0" -- required for all v1.5.0+ skills so version filters can target them.
allowed-tools: -- explicit tool list. Omit only when the skill truly needs every tool. Be explicit when you can.
model: -- only when the skill must pin a specific model (rare).
argument-hint: -- shown after the slash command in the picker.
First line above frontmatter: <!-- IJFW: narration-not-applicable --> (or narration-applicable) -- declares how the skill interacts with the narration hook.
3. Shape -- terse vs structured
Skill kind
Shape
Length
Example
One-shot trigger (single output)
Terse: frontmatter + 5-30 line body
<60 lines
ijfw-review (43 lines), ijfw-debug (52 lines)
Multi-move discipline
Structured: numbered moves with green-light gates
80-180 lines
ijfw-tdd (104 lines), this file
Reference / API doc
Structured + references/ files
header in SKILL.md, weight in references/
offload heavy content
Hot-load on trigger, unload when done. Nothing in the body should pretend to be always-loaded -- the user will only see it when the trigger fires.
4. Anti-patterns -- do not ship these
Description-as-definition. "Skill for the code review process" tells Claude what the skill IS. Claude needs to know WHEN to load it. Rewrite as "Use when reviewing code. Trigger: review, /ijfw-review".
Forbidden workflow keywords in description.step, phase, mode, workflow, stage, process, sub-skill, skill-name -- the lint will FAIL the build. Use trigger phrases instead.
Overlap with existing skills. Before writing, list claude/skills/ and read any neighbour whose name or trigger overlaps. Extend an existing skill rather than fragmenting the trigger surface.
Embedded prompts for the user to paste. The skill IS the prompt. The user just speaks the trigger -- they should never see "copy this and paste into chat".
Undeclared tool dependencies. If the skill calls Bash, Edit, or an MCP tool, either declare them in allowed-tools: or accept inherit-everything. Do not silently assume tools exist.
Documentation about the topic. A skill explains how to DO a thing, not how the thing works in the abstract. If you want to teach concepts, write a reference file under references/ and link from a skill that DOES something with the concept.
Skill that is really an agent. If the work needs an isolated context window or a different model, write an agent (claude/agents/<name>.md) and dispatch via the Agent tool. Skills are for in-context discipline; agents are for delegated work.
5. Validation flow -- always run, always pass
After writing the file:
# 1. Lint CSO discipline. Must show PASS for your new skill, no new FAILs.
bash scripts/lint/check-skill-descriptions.sh 2>&1 | grep -E "<your-skill-name>|FAIL"# 2. Trigger-test in a fresh Claude session. Open a new conversation and say# one of the verbatim trigger phrases from your description. Confirm the# skill loads and runs as expected. If it does not fire, your description# does not match how users actually speak -- rewrite and re-test.
A skill that lints PASS but does not fire on its own trigger is broken. Both checks must pass before commit.
6. Multi-domain consideration
IJFW skills should work across software, books, campaigns, and design wherever the discipline transfers. TDD applies to code AND book continuity AND campaign metrics; review applies to PRs AND chapter drafts AND landing-page copy. Write skills domain-neutral by default.
Domain-specific is rare and must be justified: e.g. ijfw-ui-spec is intentionally frontend-only because UI contracts have no analogue in book writing. When you go domain-specific, name it in the description so the trigger surface stays honest: "Use when designing a frontend feature".
7. Where skills live -- canonical layout
claude/skills/<skill-name>/
SKILL.md # required, ~140-180 lines max for structured skills
references/ # heavy reference material the skill links to
<topic>.md
prompts/ # reusable prompt fragments the skill renders
<name>.txt
templates/ # boilerplate the skill emits
<name>.template.md
Cross-platform skills also land at codex/skills/<name>/SKILL.md, gemini/extensions/ijfw/skills/<name>/SKILL.md, hermes/skills/<name>/SKILL.md, wayland/skills/<name>/SKILL.md. The lint script checks every present platform.
8. When NOT to write a new skill
You want to ...
Do this instead
Extend behaviour of an existing trigger
Edit the existing skill. Adding a second skill for the same trigger fragments routing.
Run heavy work in an isolated context
Write an agent at claude/agents/<name>.md with model: + allowed-tools: + a body that gets dispatched via the Agent tool.
Enforce a mechanical rule
Write a hook (claude/hooks/) or a lint script (scripts/lint/). Hooks are deterministic; skills are judgment calls.
Document a concept users should read
Add a doc under docs/. Skills are loaded on trigger, not for browsing.
Pin a project convention
Add it to CLAUDE.md or AGENTS.md. Skills are reusable across projects; conventions are local.
Done When
claude/skills/<name>/SKILL.md exists with correct frontmatter (name, description, since: if v1.5.0+).
Description starts with "Use when" and lists verbatim trigger phrases plus the /slash-command alias.
Description contains NO forbidden words (step, phase, mode, workflow, stage, process, sub-skill, skill-name).
Description under 1024 chars.
bash scripts/lint/check-skill-descriptions.sh shows your skill as PASS and introduces no new FAILs.
Trigger-tested in a fresh Claude session -- the skill actually fires when a user speaks a trigger phrase.
Shape matches purpose: terse (<60 lines) for one-shot, structured (80-180 lines) for multi-move discipline.
No overlap with an existing skill (or the existing skill was extended instead).
If declared since: "1.5.0", picked up by version filters.
References
Lint: scripts/lint/check-skill-descriptions.sh -- the authoritative CSO rules.
