SuperAPI AI Directory
Partner中文AI newsFavorites
Skillscreate-skill

create-skill

Agents & models

View repository

Updated 10-15 23:42

About this skill

create-skill

skills/<library>/ An orchestration skill for creating a reference skill for a new library/framework.

Usage

/create-skill zod https://zod.dev    # library と base_url を指定
/create-skill hono                   # base_url 省略時は調査で特定する
Argument Required Description
library Required Name of the library/framework to add (e.g., zod, hono）
base_url Optional Base URL of the official documentation. If omitted, it is identified in Step 1-2

main (this skill itself) focuses on planning, delegation, integration, and reporting, and delegates document reading and references writing to specialized Agents (.claude/rules/delegation.md compliant).

Determine base_url: If it was not passed as an argument, have the skill-coverage-analyzer in Step 1 report the canonical URL of the official documentation (if unknown, identify it via WebSearch in reference-researcher mode=survey at the beginning of Step 3). Record the confirmed URL in the Step 2 plan, and passthis confirmed base_url to all Agents in Step 3 and Step 4 (so that each Agent does not individually re-search for the URL).

Flow

Step 1: Check whether the skill is necessary

Delegate to **skill-coverage-analyzer (subagent_type: skill-coverage-analyzer, model: opus)** and have it check the following. Parameters to pass: target(= the argument's <library>）、base_url(only if passed as an argument).

  • Whether it overlaps with existing skills
  • Enumerate scope candidates (e.g., core, plugins, cli) from the scale and structure of the documentation
  • base_url If is unspecified, have it also report the canonical URL of the official documentation

If it is determined that the skill is unnecessary, confirm with the user and decide whether to abort.

Step 2: Create an implementation plan

create-plan Invoke the skill and_/local-plans/<library>-skill.md create a plan in .

Include the following in the plan.

  • Overview of the skill and target documentation URL
  • Scope splitting policy and the list of categories handled by each scope
  • references directory structure (tree format)
  • Implementation steps and completion criteria
  • Verification method

If the scope was not finalized in Step 1, treat the scope split in the plan as provisional, finalize it based on the mode=survey survey results at the beginning of Step 3, and then reflect it in the plan (update the scope splitting policy after the survey).

Step 3: Generate references in parallel

Following the scope split finalized in Step 1 and Step 2, for each scope reference-researcher（subagent_type: reference-researcher、model: sonnet）をin parallellaunch, and and the references for the target categories (individual page .md) README.md Generate the index. Each reference-researcher obtains the official documentation navigation on its own and finalizes the page list (main does not read the official documentation).

If the scope is not yet determined, first launch one reference-researcher with mode=survey to investigate only the site structure (survey mode does not write out files and reports only category candidates. In this launch, do not pass scope・output_dir ). Based on the result (category list), split the scope, then launch each scope in parallel with mode=full(default). If there are many categories, split with the guideline of 1 Agent = 1 category.

Parameters to pass to each Agent:

Parameter Content
library Library name
base_url URL of the official documentation finalized in Steps 1–2 (including the one identified when omitted)
scope Assigned scope. In principle, 1 Agent = 1 category. Specify comma-separated only when combining multiple categories
output_dir For a single category, skills/<library>/references/<category>/. For multiple categories (comma-separated scope), pass skills/<library>/references/(root), and the Agent creates subdirectories per category
mode full(default). Only the preliminary investigation when the scope is not yet determined is launched with survey launch with

Step 4: Generate samples and scripts in parallel

In parallel with the generation of references, or after completion, launch the following in parallel.

  • sample-curator（subagent_type: sample-curator、model: sonnet） — skills/<library>/samples/ Generate examples of typical use cases. Parameters to pass: library, skill_dir（skills/<library>/）, base_url(finalized in Steps 1–2)
  • script-collector（subagent_type: script-collector、model: sonnet） — skills/<library>/scripts/ Generate a collection of copy-pasteable commands. Parameters to pass: library, skill_dir（skills/<library>/）, base_url(finalized in Steps 1–2)

If, due to the nature of the library, samples or scripts are judged unnecessary (e.g., guideline-type skills), they may be skipped.

Step 5: Create SKILL.md

Step 3（references）と After confirming the completion of the Agents actually launched in Step 4(samples / scripts), delegate to **skill-author (subagent_type: skill-author, model: sonnet)** to create skills/<library>/SKILL.md . If samples / scripts were judged unnecessary in Step 4 and skipped, proceed without waiting for their completion, using only the completion of Step 3 (do not wait for the completion of Agents that were not launched). By finalizing samples / scripts first, they are reliably reflected in the SKILL.md tree and mapping table.

Parameters to pass:

Parameter Content
library Library name
skill_dir skills/<library>/
user-invocable Distributed skills are normally false(skill-anatomy compliant)

Content generated by skill-author:

  • YAML frontmatter（name / description / user-invocable）
  • Directory tree
  • Exploration procedure
  • Task → category → README mapping table

skill-author does not optimize the description itself; if optimization is needed, it returns a "delegation proposal" to main.If such a proposal is made, or if the reference-linter in the subsequent Step 6 points out a deficiency in the description, main delegates to description-optimizer (subagent_type: description-optimizer, model: sonnet)(Parameters to pass: skill_dir＝skills/<library>/）。

Step 6: Verify quality

Launch the following Agents to verify quality. In all cases, limit the verification target with skill_dir（skills/<library>/) (if unspecified, the linter scans all of skills/ and the validator's target directory becomes indeterminate).

  • reference-linter（subagent_type: reference-linter、model: haiku） — Have it check that the references body, samples/scripts body, each README index, and SKILL.md frontmatter conform to the format (parameters to pass: skill_dir）
  • skill-structure-validator（subagent_type: skill-structure-validator、model: haiku） — Have it check whether SKILL.md and the directory structure conform to skill-anatomy (parameters to pass: skill_dir）

Return the pointed-out issues to the relevant Agent according to their content, and re-verify after correction.

  • Deficiency in references body → reference-researcher
  • Deficiency in samples body → sample-curator
  • Deficiency in scripts body → script-collector
  • Each category README.md Deficiency in index → readme-indexer
  • SKILL.md の description(hit rate, length, YAML # -caused deficiencies) → description-optimizer
  • SKILL.md tree, mapping table,description Deficiencies in frontmatter other than → skill-author

Step 7: Update CLAUDE.md

update-docs Invoke the skill and reflect the new skill in the "Current Skills" list and repository structure tree in CLAUDE.md.

Step 8: Commit (optional)

If the user instructs to commit, create-commit Create a commit with the skill.

Verification

  • ls skills/<library>/SKILL.md Confirm that the file exists with
  • head -10 skills/<library>/SKILL.md Confirm that frontmatter (name / description / user-invocable) is correctly written with
  • ls skills/<library>/references/ Confirm that each category's references and README.md exist under
  • If samples / scripts were generated, ls skills/<library>/samples/ ls skills/<library>/scripts/ confirm the existence of; if skipped, confirm that the relevant row is not included in the SKILL.md tree/mapping table
  • Confirm that the reference-linter and skill-structure-validator verifications are PASS.
  • Confirm that <library> has been added to the "Current Skills" list in CLAUDE.md

Notes

  • .claude/ When editing under .claude/rules/dotclaude-via-temp.md follow _/dotclaude/ and do so via
  • The body of references is written in the language of the original document (usually English).
  • The description and Japanese explanatory text in SKILL.md follow .claude/rules/description-style.md and .claude/rules/japanese-style.md comply with
  • main itself does not directly read the official documentation, and always delegates to a specialized Agent.

Related

Agents: reference-researcher, sample-curator, script-collector, skill-author, description-optimizer, reference-linter, skill-structure-validator, skill-coverage-analyzer

Rules: delegation, skill-anatomy, reference-template, description-style

Skills: create-plan, update-docs, create-commit