How to use the Skillify Meta Skill agent skill

What does the Skillify skill make the agent do?

Skillify tells the agent how to turn a raw feature into a properly structured skill. It walks through a 15-item checklist covering the skill's instruction file (SKILL.md), any deterministic script, unit and integration tests, LLM evals, registration in the skill resolver, and plugin membership. Items are numbered 0 and 1 through 14, with 3b sitting alongside item 3.

When a skill already exists, skillify improves it instead of rewriting it. It reads the existing files, identifies what changed (a bug report, user feedback, new input or a quality gap), edits the existing files, and re-runs the checklist touching only items that fail.

Skillify also sets a no-regression rule: any edit to a skill must score at least as well as the previous iteration on the same cross-modal evaluation. Each skill declares an eval contract with its goal, skill-specific dimensions and hard-fail conditions, so the evaluation judges the skill against its own purpose.

Install the Skillify skill

Download the skillify ZIP from this page. Unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only in that project. The package contains a single file, SKILL.md.

If you use the Claude apps instead, upload the ZIP under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: upload the ZIP under Settings → Capabilities → Skills

How do you trigger the skill with natural requests?

You do not need a special command. Ask in plain language using a phrase the skill is written to respond to. The listed trigger phrases include "skillify this", "skillify", "is this a skill?", "make this proper", "add tests and evals for this", "check skill completeness", "run skillify on a skill" and "did this skill regress".

• "Is this a skill?" before you scaffold anything, to check whether the work deserves skill infrastructure
• "Skillify this" on a feature you have just built
• "Check skill completeness" to audit an existing skill against the checklist
• "Did this skill regress?" after editing a skill

What does a typical run look like?

For a new feature, the skill first checks whether it qualifies as a skill. It asks whether the task will be invoked two or more times, whether there is more than 20 lines of logic, and whether a user would naturally phrase a trigger for it. If any answer is no, the skill says the work should be a plain script and stops. It also checks for overlapping trigger phrases in the resolver before creating a near-duplicate.

For an existing skill, the run reads SKILL.md, the script, the routing eval file and the test file, identifies the delta, and improves those files in place. Each change is checked against the no-regression rule before it ships.

• Phase 0 decides new skill versus improvement
• Cross-modal evaluation runs before tests, so tests lock in behavior that has already been judged good
• Scheduled-job skills must re-run a representative task and evaluate it before an edit ships

What do you need before using Skillify?

Skillify assumes a working repository and some tooling. The instructions expect a skills/RESOLVER.md file and either a skills/plugin-exclusions.json or an openclaw.plugin.json file. They also reference the gbrain CLI with its skillify check and eval cross-modal commands. The cross-modal evaluation step needs access to three frontier models from three different providers.

The skill runs scripts and uses the network, so review what it will execute before running it on your machine.

• A repository containing skills/RESOLVER.md and skills/plugin-exclusions.json or openclaw.plugin.json
• The gbrain CLI with the skillify check and eval cross-modal commands
• Access to three frontier models from three different providers for the cross-modal step

Safety notes: scripts and network access

The skill is marked as running scripts and using the network. Its frontmatter lists the tools exec, read, write and edit, and it is marked as mutating, meaning it changes files. Run it inside a version-controlled repository so every change can be reviewed and reverted.

The cross-modal step sends skill output to three different model providers for critique. If your skill content is sensitive, check that this is acceptable before running the evaluation.

• Use a version-controlled repository so edits are reviewable
• Confirm that sending skill output to external model providers is acceptable
• The skill includes hard-fail rules, including no edits to scheduled skills without re-running a representative task

Tips for getting good results

Give skillify a specific feature or a specific skill rather than a broad area. The skill itself says that a target spanning several separate intents should be split into separate skillify targets. Keep the eval contract concrete, since generic dimensions make the cross-modal judgment less useful.

Use the checklist numbers when you talk about gaps. Asking to run skillify items 4 through 6 targets only those parts of the checklist.

• Name one capability per run
• Keep the eval contract goal concrete and skill-specific
• Refer to checklist items by number when asking for targeted fixes