Agent Skills Creator

Agent Skills Creator

Create and improve skills in the Agent Skills open format: full lifecycle from pattern selection through validation and README update.

• IS: creating new agent skills and auditing or rewriting existing ones: SKILL.md, references, rules folders, scripts, evaluations.
• IS NOT: AGENTS.md/CLAUDE.md instruction files (use agents-md) or general documentation quality (use docs-writing).

Choose a Mode

• New skill → Creation Workflow below.
• Audit, improve, or rewrite an existing skill → references/improving-existing-skills.md, which scores eleven audit dimensions, runs an ordered rewrite, then reuses Steps 5-8 for validation and shipping.
• Simplify a skill or update a collection → same reference, plus references/capability-delta.md for retention decisions and the collection ledger.
• Skill never triggers, triggers on the wrong prompts, or its description is being trimmed in a crowded listing → the Routing Evals section of references/evaluation-and-iteration.md, then the description. Body edits do not fix routing.
• Skill stops after a first draft and hands back for review when nobody asked → "Say Where the Work Ends" in references/authoring-tips.md; the fix is the scope of done and the review checkpoints, not the steps.

Reference Files

The validator is the single local statement of mechanical gates. From the repository root:

bashCopy code
skills/agent-skills-creator/scripts/validate.sh skills/<name>   # one skill
skills/agent-skills-creator/scripts/validate.sh --all           # every skill in the repo

Creation Workflow

Copy this checklist to track progress:

textCopy code
Skill creation progress:
- [ ] Step 1: Choose a pattern
- [ ] Step 2: Create directory and frontmatter
- [ ] Step 3: Write SKILL.md body
- [ ] Step 4: Add reference or rule files
- [ ] Step 5: Validate
- [ ] Step 6: Update README.md (and docs/skills.mdx where the repo has one)
- [ ] Step 7: Smoke-test installation
- [ ] Step 8: Evaluate and iterate

Step 1: Choose a pattern

Simple/hub, workflow, rules-based, or mixed. references/skill-patterns.md has the shapes, the in-repo example for each, and the problem-to-pattern affinity table.

Step 2: Create directory and frontmatter

Create skills/<name>/SKILL.md with name and description, the --- on line 1. Write the description as a model trigger, not a human summary: what it does, what it covers, then "Use when..." with the phrases users actually say, and the key use case first because the listing trims descriptions from the tail when it runs over budget. Keep it as short as it can be while still saying when it applies: every installed description sits in every session, and one that over-claims loads a skill that does not help. validate.sh enforces the limits, so write for routing and let the script police the constraints.

Decide where the skill will run before adding any other field. Use portable fields for this collection; put genuine runtime prerequisites in compatibility. Host-specific fields require current host documentation and a deliberately host-specific package. State authorization boundaries in the body when a workflow deploys, sends, or spends, and grant the safe loops just as explicitly; the body carries both sides of that envelope. The format reference distinguishes these contracts.

Step 3: Write SKILL.md body

references/authoring-tips.md carries the judgement. Apply:

• Open with an IS/IS-NOT pair when adjacent skills exist or scope creep is likely ("Open with Boundaries")
• Add only context the agent lacks ("Don't State the Obvious"); use consistent terminology
• Phrase guidance as an outcome, reserving absolutes for safety, data loss, format contracts, and observed failures ("Judgement Over Rules")
• Check nothing here contradicts the harness, a sibling skill, or the repo AGENTS.md; route instead of restate ("Don't Fight the Harness or a Sibling")
• Keep the opinions that make the skill worth invoking; cut only what the target agent already does unprompted ("Cut Constraints, Keep Opinions"). On current frontier models over-prescription is not merely wasted tokens: instructions carried forward from older models are often too prescriptive and lower output quality, so the constraint cut is correctness work
• Match degrees of freedom to fragility: prose for open-ended work, exact commands for fragile or destructive ops ("Degrees of Freedom")
• Reach for named content patterns: template for fixed output, examples only where style is the deliverable, conditional for decision points
• Write the permission side of the envelope, not only the restriction: grant a known-safe loop with the reason it is safe, and keep confirmation for what deploys, sends, spends, or writes outside the working tree ("Grant Permission, Don't Just Restrict")
• State the workflow dependencies, what the finished state includes, and the completion evidence; scope and evidence are different sentences ("Say Where the Work Ends"). Keep a stop-for-review step only where the decision is genuinely the user's; current frontier models honor a mandatory pause literally and end the task there. Add a checklist only when it helps track a long or resumable task
• Put the deliverable, routing, and task-specific constraints first; loading and compaction behavior depend on the host. A skill with several workflows makes SKILL.md a router (mode choice, shared contract, one pointer per workflow) and keeps each workflow's body in its own reference ("Use the File System for Progressive Disclosure")
• Build a Gotchas section from observed failures: the highest-signal content in any skill

Step 4: Add reference or rule files

• Workflow/mixed: a references/ folder, each file linked from SKILL.md with a "Read when" condition
• Rules-based: a rules/ folder; references/rules-folder-structure.md covers _sections.md, _template.md, and file naming
• Simple/hub: track files alongside SKILL.md, linked from a tracks table

Prefer a reference that is code. An existing implementation or a test suite pins a contract better than prose describing it ("Reference-as-Spec"). Split by loading condition, not line count: two topics read at different moments are two files.

Advanced, all covered in references/executable-code.md and references/format-specification.md: scripts/ for executables the agent composes, resolved relative to the installed SKILL.md; Claude Code can substitute ${CLAUDE_SKILL_DIR}; !`command` injection for data the skill always needs at invocation (a diff, PR comments); hooks frontmatter for a PreToolUse gate that should exist only while the skill is active; config.json for setup context that would otherwise be re-asked every session.

Step 5: Validate

bashCopy code
skills/agent-skills-creator/scripts/validate.sh skills/<name>

Output separates format requirements from house conventions, including recommendations adopted as local gates. Fix every FAIL. Run skills-ref validate <dir> for an independent metadata and naming check. Neither check establishes behavioral quality or compatibility with every host.

Step 6: Update README.md

Add a bullet under the matching category heading, and bump the skill count near the top of the README:

markdownCopy code
- **[<skill-name>](./skills/<skill-name>/SKILL.md)**: <one-line description>

Categories: Architecture, Design, Writing, Quality, Shipping, Authoring. validate.sh verifies the bullet and the count. A repo that also ships docs/skills.mdx needs the same bullet there; the check is conditional on that file existing, so it stays silent in a repo without one.

Step 7: Smoke-test

When installation behavior changed, install the edited local source into a disposable target using the installer's documented local-source options. Verify the loaded SKILL.md and a bundled reference or script against the working copy. Do not overwrite a global installation to smoke-test an edit, and do not install the remote default branch as evidence for unpushed changes.

Step 8: Evaluate and iterate

references/evaluation-and-iteration.md. Write 2-3 scenarios in evals/evals.json, add assertions after the first run, and measure with-skill against without-skill in fresh sessions. Test routing separately with should-trigger and near-miss prompts. Test on each target model, and ablate any rule you suspect is dead weight: delete it, rerun the scenarios, keep it only if one regresses.

Gotchas

• The installed copy under ~/.agents/skills/<name>/ is a copy, not a link to your repo. Editing the repo changes nothing in a running session, and the stale copy loads silently, so a skill can be several commits behind while appearing correct. Verify that an evaluation loads the edited local source.
• A reference-chain failure calls for moving the load condition into SKILL.md, not disguising the same dependency with different wording.
• toc-over-100-lines wants ## Contents inside the first 20 lines of any reference over 100 lines. A TOC further down does not count, and the file fails while looking fine.
• readme-skill-count compares the README's stated count against find skills -maxdepth 2 -name SKILL.md. Adding a reference file to an existing skill does not change it; only adding or removing a skill does.
• --agent on skills add is variadic and space-separated (--agent codex cursor); it consumes arguments until the next one starting with -. A comma-separated list is validated element-wise and rejected whole as one invalid name. Do not conclude from an empty ~/.codex/skills that the install failed: any agent whose skillsDir is .agents/skills is treated as universal and installed to ~/.agents/skills/, which those agents read directly.
• A description that omits "Use when" fails description-triggers outright, but a description that has the phrase and the wrong trigger words fails nothing and simply never routes. The validator cannot see this; only a routing eval can.
• description-length passes anything under 1024 characters, and a collection of descriptions that each pass can still overrun the host's listing budget together, at which point the host shortens every one of them and the "Use when" clauses go first. Judge a description in the listing next to its siblings, not alone.
• A step reading "stop here and present the result for review" ends the task on current frontier models even when the user asked for the whole job and is not watching. The step reads as a safety rule, so it is rarely questioned during a constraint cut, and the skill ships stopping one implementation short.

Anti-patterns

• Usage examples standing in for an expressive interface; name the parameters and enums instead
• A rule stated in SKILL.md and again in a script's --help, a tool description, or a rule file
• Dumping the full specification into the SKILL.md body instead of a reference file
• Time-sensitive content ("before August 2025, use..."), including model names as the reason a rule exists
• A context: fork skill whose body is guidelines rather than a task; the subagent gets conventions and no prompt, and returns nothing
• Vague names (helper, utils, tools, documents, data) that give the model nothing to route on
• Magic numbers in scripts with no justifying comment
• Shipping without testing across the capability tiers and effort levels the skill will actually run under; what reads well to a frontier model may underspecify a small fast one, and a step the model only volunteers at high effort is a step the workflow does not really have

Related Skills

• agents-md for auditing AGENTS.md/CLAUDE.md instruction files
• docs-writing for documentation quality rules

Maintenance only: evals/evals.json holds the behavioural scenarios and routing prompts for anyone changing this skill. It never loads during a user task, which is the baseline Phase A of references/improving-existing-skills.md asks for.