create-skill
**WORKFLOW SKILL** - Create and register a complete portable agent skill in a managed skills-repo.config.json repository or an existing skills repository. Produces focused instructions, documentation, deterministic tests, Vally evals, locked tooling, registration, and optional va
Install
npx skills add https://github.com/jongio/skills/tree/main/skills/create-skill
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install jongio-skills@llmmart
git clone https://github.com/jongio/skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole jongio/skills collection as a plugin from our marketplace. Git is the plain clone.
README
create-skill
Create and register a complete portable agent skill without missing tests, evals, catalog entries, or thumbnail provenance.
Install
npx skills add jongio/skills --skill create-skill
Reload skills, then invoke:
/create-skill release-notes-helper
/create-skill art release-notes-helper
/create-skill check release-notes-helper
Repository generators can use the stable fixture contract:
node scripts/create-skill.mjs fixture --input fixture.json --repo-root ./target --dry-run
node scripts/create-skill.mjs fixture --input fixture.json --repo-root ./target --approve HASH
Fixture mode is noninteractive and accepts the documented version 1 payload in references/fixture-mode.md.
Guarantees
- Managed
skills-repo.config.jsonpaths are authoritative. - Existing repository conventions are discovered and ambiguous registration fails closed.
- Every mutation has a zero-write dry-run. Art actions use a persisted single-use approval token.
- Azure OpenAI and OpenAI use exact origins, disabled redirects, and zero retries.
- Custom art may use any approved provider and delivery workflow.
- Every image passes complete PNG, CRC, raster, metadata, dimension, and size validation.
- Skill thumbnails are the tracked source; catalog builds generate public copies from frontmatter.
- Prompt and non-secret provenance are recorded when the repository supports that surface.
Development
New skills include metadata.version: "1.0.0" in SKILL.md. Registration
preserves that version (or the existing marketplace version for legacy skills).
In repositories with scripts/version-skills.mjs, use its bump command for
releases and check before publishing. Skill releases are separate from
private npm development-tooling versions.
npm ci --ignore-scripts
npm test
npm run eval:lint
Tests are deterministic. They use temporary repositories, injected clients, and local PNG buffers. They never perform a real network request or billed image generation.
Generated skills pin @microsoft/vally-cli to 0.20.0 and reuse a matching
repository Vally tool lockfile. They carry a scoped Copilot SDK override to Zod
4.6.7. The source lock must already resolve that override; otherwise generation
stops before writing files. Update older repository tooling before authoring.
License
MIT. See LICENSE.
Skill manifest
Caution
Read and follow the shared safety rules before any action. Preview local writes and obtain explicit approval. Git and GitHub writes require their own approval and are outside this skill.
Create Skill
Create one complete skill and register every repository surface without guessing. The workflow plans first, fails closed on ambiguous conventions, and applies one approved change set.
Commands
| Command | Purpose |
|---|---|
create-skill <name> |
Create and register a new skill with deterministic placeholder art. |
create-skill art <name> |
Preview and create or ingest thumbnail art for an existing skill. |
create-skill register <name> |
Preview and register an existing complete skill without replacing its authored files. |
create-skill check <name> |
Validate the skill shape and thumbnail parity without writing. |
create-skill fixture --input <json> --repo-root <path> |
Deterministically create a validated fixture skill without interview prompts. |
Every mutating command starts with --dry-run. Apply the unchanged plan only after the user
approves its single-use hash.
Workflow
1. Resolve the request
Collect only missing values:
- Skill name in lowercase-dash form.
- A one-sentence summary.
- Specific
USE FORphrases that users will type. - Specific
DO NOT USE FORboundaries and the sibling workflow that should win. - Copyright holder, when repository conventions do not identify one.
- Art choice: Azure OpenAI GPT Image 2, OpenAI Images, custom workflow, or placeholder.
Search installed and repository skills for overlapping triggers before creation. If another skill already owns the same job, stop and propose extending it.
2. Discover the repository
Run from the target repository, not this installed skill directory.
- Prefer the nearest
skills-repo.config.json. - Validate managed identity and settings before deriving canonical repository paths.
- Without managed config, discover the existing populated
skills/directory and known registration surfaces. - Stop before any write when multiple skill roots or catalog surfaces are plausible.
See repository discovery for the exact contract.
3. Preview creation
Invoke this script by absolute path while keeping the target repository as the working directory:
node <skill-dir>/scripts/create-skill.mjs <name> \
--summary "<summary>" \
--use-for "<trigger phrases>" \
--do-not-use-for "<boundaries>" \
--author "<holder>" \
--dry-run
Show the complete JSON preview and approval token. The preview includes every skill file and every registration surface. Ask the user to approve that exact hash once.
4. Apply creation
Repeat the same command and replace --dry-run with --approve <hash>. Do not alter any input
between preview and apply. The tool rejects a stale hash, existing skill directory, symlinked
destination, missing locked Vally toolchain, ambiguous registration, or invalid PNG.
Creation includes:
SKILL.mdwithname,description, andmetadata.versionfrontmatter.README.md,LICENSE,package.json, andpackage-lock.json..vally.yaml, deterministic tests, andevals/<name>/eval.yaml.thumbnail.pngas the tracked source for catalog synchronization when a catalog is present.- Marketplace, plugin, README, eval workflow, Dependabot, and catalog registration when present.
- Prompt and non-secret provenance in
docs/thumbnail-prompts.mdwhen present.
5. Choose art
The initial placeholder is deliberate and deterministic. Replace it only when requested.
Built-in providers
Preview the exact action:
node <skill-dir>/scripts/create-skill.mjs art <name> \
--provider azure-openai \
--prompt "<exact prompt>" \
--dry-run
Supported choices:
azure-openai: Azure OpenAI GPT Image 2 at the exact configuredhttps://<resource>.openai.azure.comorigin.openai: OpenAI Images at the exacthttps://api.openai.comorigin.placeholder: deterministic local PNG with no network or billed call.
The placeholder and check commands run on bare Node. Before a network provider, run
npm ci --ignore-scripts in the installed create-skill directory so its locked SDKs are
available.
Show provider, exact endpoint, model, prompt, targets, billing status, one attempt, and no fallback.
Ask for approval of the single-use action hash. Then repeat with --approve <hash>.
Built-in requests disable redirects and SDK retries. Never retry automatically. Never switch providers after a timeout, redirect, rate limit, service error, invalid response, or rejected PNG. Report the failure and ask the user what to do.
Custom provider or delivery workflow
Ask: "Describe any provider and delivery workflow you want me to use for the thumbnail."
The answer is open-ended. It may name an API, MCP tool, browser flow, local application, uploaded attachment, shared storage location, human handoff, or another delivery mechanism. Never reduce the choice to a local command.
Convert the answer into an exact action preview containing:
- Provider and destination.
- Exact prompt and non-secret data sent.
- Exact tool calls, API origin, browser actions, or human handoff steps.
- Billing or external side effects.
- One attempt and no fallback.
- How the final PNG will arrive in the target repository workspace.
- The skill and catalog paths that will receive the validated bytes.
Show that preview and obtain explicit single-use approval before taking the custom action. Execute only the approved steps once. Never place credentials in the preview or provenance.
After delivery, save the result as a repository-relative regular file and ingest it:
node <skill-dir>/scripts/create-skill.mjs art <name> \
--provider custom \
--custom-description "<approved provider and delivery steps>" \
--custom-provider "<provider name>" \
--custom-model "<model or method>" \
--delivery "<delivery method>" \
--input "<repository-relative PNG>" \
--dry-run
Approve the displayed hash, then repeat with --approve <hash>. The CLI never executes a custom
command. It accepts the delivered PNG only after the same strict validation used for built-in art.
See art and provenance.
6. Validate
Run:
node <skill-dir>/scripts/create-skill.mjs check <name>
npm test
npm run eval:lint
Run devx skills doctor, strict skill checking, routing, and overlap checks when those tools are
available in the target repository. Do not run the full capability eval unless the user accepts
its agent usage.
7. Generate a deterministic fixture
Repository generators may create functional example skills through the stable fixture contract:
node <skill-dir>/scripts/create-skill.mjs fixture \
--input "<fixture.json>" \
--repo-root "<target-repository>" \
--dry-run
Parse the JSON preview, obtain or propagate approval for its hash, then repeat the unchanged
command with --approve <hash>. Fixture mode is always noninteractive. It accepts only the version
1 schema, deterministic built-in placeholder art, and the normal repository registration path.
It never bypasses dry-run, approval, validation, atomic writes, or provenance.
Read fixture mode for the exact payload and exit contract.
Repository generators use create-skill register create-skill --dry-run, followed by the
unchanged command with --approve <hash>, after catalog staging. This registers the bundled
authoring skill through the same atomic path without rewriting its skill files.
Safety Invariants
- Dry-run never writes files or calls an image provider.
- Approval binds the exact inputs and target paths. Changed inputs require a new preview.
- Provider requests make one attempt with no retry, redirect, or fallback.
- Custom workflows require an exact preview and approval before any external action.
- Provider output is untrusted until the complete PNG validator accepts it.
- The validated skill thumbnail is the tracked source; catalog builds generate public copies.
- Secrets never enter prompts, previews, logs, generated files, or provenance.
- Existing files are preserved when planning fails. Apply uses staged sibling files and rollback.
- No git staging, commit, push, issue, pull request, or repository setting change is part of this skill.
Exit Criteria
New skills declare
metadata.versionas a quoted SemVer string. Preserve existing skill versions during registration; use the repository's version helper to bump a release and synchronize plugin listings when available.The skill has every required file and only portable SKILL.md frontmatter keys.
Every discovered or configured registration surface includes the skill.
The tracked skill thumbnail passes strict PNG validation; catalog builds generate the public copy.
Prompt and non-secret provenance are recorded when that repository surface exists.
Deterministic tests and Vally lint pass.
The final report lists exact changed files, commands, results, and any blocked work.
Files (skills)
-
evals
-
create-skill
-
eval.yaml 5.1 KB
Loading…
-
-
README.md 289 B
Loading…
-
-
references
-
art-and-provenance.md 2.2 KB
Loading… -
fixture-mode.md 2.1 KB
Loading… -
repository-discovery.md 2.5 KB
Loading…
-
-
scripts
-
art.mjs 6.7 KB · in bundle
-
create-skill.mjs 12.3 KB · in bundle
-
png.mjs 9 KB · in bundle
-
providers.mjs 5.1 KB · in bundle
-
registration.mjs 21 KB · in bundle
-
render.mjs 10.8 KB · in bundle
-
repository.mjs 9.4 KB · in bundle
-
versioning.mjs 6.7 KB · in bundle
-
versions.mjs 4 KB · in bundle
-
-
test
-
art.test.mjs 5.6 KB · in bundle
-
cli.test.mjs 10.8 KB · in bundle
-
helpers.mjs 4.1 KB · in bundle
-
png.test.mjs 7 KB · in bundle
-
providers.test.mjs 6.9 KB · in bundle
-
registration.test.mjs 14.4 KB · in bundle
-
render.test.mjs 6.1 KB · in bundle
-
repository.test.mjs 5.6 KB · in bundle
-
versions.test.mjs 11.7 KB · in bundle
-
-
.gitignore 29 B · in bundle
-
.npmrc 18 B · in bundle
-
.vally.yaml 65 B
Loading… -
LICENSE 1 KB · in bundle
-
package-lock.json 125.3 KB
Loading… -
package.json 1012 B
Loading… -
README.md 2.4 KB
# create-skill Create and register a complete portable agent skill without missing tests, evals, catalog entries, or thumbnail provenance. ## Install ```sh npx skills add jongio/skills --skill create-skill ``` Reload skills, then invoke: ```text /create-skill release-notes-helper /create-skill art release-notes-helper /create-skill check release-notes-helper ``` Repository generators can use the stable fixture contract: ```sh node scripts/create-skill.mjs fixture --input fixture.json --repo-root ./target --dry-run node scripts/create-skill.mjs fixture --input fixture.json --repo-root ./target --approve HASH ``` Fixture mode is noninteractive and accepts the documented version 1 payload in [references/fixture-mode.md](references/fixture-mode.md). ## Guarantees - Managed `skills-repo.config.json` paths are authoritative. - Existing repository conventions are discovered and ambiguous registration fails closed. - Every mutation has a zero-write dry-run. Art actions use a persisted single-use approval token. - Azure OpenAI and OpenAI use exact origins, disabled redirects, and zero retries. - Custom art may use any approved provider and delivery workflow. - Every image passes complete PNG, CRC, raster, metadata, dimension, and size validation. - Skill thumbnails are the tracked source; catalog builds generate public copies from frontmatter. - Prompt and non-secret provenance are recorded when the repository supports that surface. ## Development New skills include `metadata.version: "1.0.0"` in `SKILL.md`. Registration preserves that version (or the existing marketplace version for legacy skills). In repositories with `scripts/version-skills.mjs`, use its `bump` command for releases and `check` before publishing. Skill releases are separate from private npm development-tooling versions. ```sh npm ci --ignore-scripts npm test npm run eval:lint ``` Tests are deterministic. They use temporary repositories, injected clients, and local PNG buffers. They never perform a real network request or billed image generation. Generated skills pin `@microsoft/vally-cli` to `0.20.0` and reuse a matching repository Vally tool lockfile. They carry a scoped Copilot SDK override to Zod `4.6.7`. The source lock must already resolve that override; otherwise generation stops before writing files. Update older repository tooling before authoring. ## License MIT. See [LICENSE](LICENSE). -
SKILL.md 9.8 KB
--- name: create-skill description: >- **WORKFLOW SKILL** - Create and register a complete portable agent skill in a managed skills-repo.config.json repository or an existing skills repository. Produces focused instructions, documentation, deterministic tests, Vally evals, locked tooling, registration, and optional validated thumbnail art. USE FOR: create-skill, create a skill, add an agent skill, scaffold a portable skill, register a skill, add skill to managed repo, create skill in existing repo, generate an example skill from a fixture, add skill thumbnail art. DO NOT USE FOR: scaffolding a new marketplace collection from scratch (use create-skills-repo), editing one existing skill instruction without registration work (use skill-authoring), creating an agent persona. metadata: version: "1.1.1" --- > [!CAUTION] > Read and follow the shared safety rules before any action. Preview local writes and obtain > explicit approval. Git and GitHub writes require their own approval and are outside this skill. # Create Skill Create one complete skill and register every repository surface without guessing. The workflow plans first, fails closed on ambiguous conventions, and applies one approved change set. ## Commands | Command | Purpose | | --- | --- | | `create-skill <name>` | Create and register a new skill with deterministic placeholder art. | | `create-skill art <name>` | Preview and create or ingest thumbnail art for an existing skill. | | `create-skill register <name>` | Preview and register an existing complete skill without replacing its authored files. | | `create-skill check <name>` | Validate the skill shape and thumbnail parity without writing. | | `create-skill fixture --input <json> --repo-root <path>` | Deterministically create a validated fixture skill without interview prompts. | Every mutating command starts with `--dry-run`. Apply the unchanged plan only after the user approves its single-use hash. ## Workflow ### 1. Resolve the request Collect only missing values: 1. Skill name in lowercase-dash form. 2. A one-sentence summary. 3. Specific `USE FOR` phrases that users will type. 4. Specific `DO NOT USE FOR` boundaries and the sibling workflow that should win. 5. Copyright holder, when repository conventions do not identify one. 6. Art choice: Azure OpenAI GPT Image 2, OpenAI Images, custom workflow, or placeholder. Search installed and repository skills for overlapping triggers before creation. If another skill already owns the same job, stop and propose extending it. ### 2. Discover the repository Run from the target repository, not this installed skill directory. 1. Prefer the nearest `skills-repo.config.json`. 2. Validate managed identity and settings before deriving canonical repository paths. 3. Without managed config, discover the existing populated `skills/` directory and known registration surfaces. 4. Stop before any write when multiple skill roots or catalog surfaces are plausible. See [repository discovery](references/repository-discovery.md) for the exact contract. ### 3. Preview creation Invoke this script by absolute path while keeping the target repository as the working directory: ```text node <skill-dir>/scripts/create-skill.mjs <name> \ --summary "<summary>" \ --use-for "<trigger phrases>" \ --do-not-use-for "<boundaries>" \ --author "<holder>" \ --dry-run ``` Show the complete JSON preview and approval token. The preview includes every skill file and every registration surface. Ask the user to approve that exact hash once. ### 4. Apply creation Repeat the same command and replace `--dry-run` with `--approve <hash>`. Do not alter any input between preview and apply. The tool rejects a stale hash, existing skill directory, symlinked destination, missing locked Vally toolchain, ambiguous registration, or invalid PNG. Creation includes: - `SKILL.md` with `name`, `description`, and `metadata.version` frontmatter. - `README.md`, `LICENSE`, `package.json`, and `package-lock.json`. - `.vally.yaml`, deterministic tests, and `evals/<name>/eval.yaml`. - `thumbnail.png` as the tracked source for catalog synchronization when a catalog is present. - Marketplace, plugin, README, eval workflow, Dependabot, and catalog registration when present. - Prompt and non-secret provenance in `docs/thumbnail-prompts.md` when present. ### 5. Choose art The initial placeholder is deliberate and deterministic. Replace it only when requested. #### Built-in providers Preview the exact action: ```text node <skill-dir>/scripts/create-skill.mjs art <name> \ --provider azure-openai \ --prompt "<exact prompt>" \ --dry-run ``` Supported choices: - `azure-openai`: Azure OpenAI GPT Image 2 at the exact configured `https://<resource>.openai.azure.com` origin. - `openai`: OpenAI Images at the exact `https://api.openai.com` origin. - `placeholder`: deterministic local PNG with no network or billed call. The placeholder and check commands run on bare Node. Before a network provider, run `npm ci --ignore-scripts` in the installed `create-skill` directory so its locked SDKs are available. Show provider, exact endpoint, model, prompt, targets, billing status, one attempt, and no fallback. Ask for approval of the single-use action hash. Then repeat with `--approve <hash>`. Built-in requests disable redirects and SDK retries. Never retry automatically. Never switch providers after a timeout, redirect, rate limit, service error, invalid response, or rejected PNG. Report the failure and ask the user what to do. #### Custom provider or delivery workflow Ask: "Describe any provider and delivery workflow you want me to use for the thumbnail." The answer is open-ended. It may name an API, MCP tool, browser flow, local application, uploaded attachment, shared storage location, human handoff, or another delivery mechanism. Never reduce the choice to a local command. Convert the answer into an exact action preview containing: 1. Provider and destination. 2. Exact prompt and non-secret data sent. 3. Exact tool calls, API origin, browser actions, or human handoff steps. 4. Billing or external side effects. 5. One attempt and no fallback. 6. How the final PNG will arrive in the target repository workspace. 7. The skill and catalog paths that will receive the validated bytes. Show that preview and obtain explicit single-use approval before taking the custom action. Execute only the approved steps once. Never place credentials in the preview or provenance. After delivery, save the result as a repository-relative regular file and ingest it: ```text node <skill-dir>/scripts/create-skill.mjs art <name> \ --provider custom \ --custom-description "<approved provider and delivery steps>" \ --custom-provider "<provider name>" \ --custom-model "<model or method>" \ --delivery "<delivery method>" \ --input "<repository-relative PNG>" \ --dry-run ``` Approve the displayed hash, then repeat with `--approve <hash>`. The CLI never executes a custom command. It accepts the delivered PNG only after the same strict validation used for built-in art. See [art and provenance](references/art-and-provenance.md). ### 6. Validate Run: ```text node <skill-dir>/scripts/create-skill.mjs check <name> npm test npm run eval:lint ``` Run `devx skills doctor`, strict skill checking, routing, and overlap checks when those tools are available in the target repository. Do not run the full capability eval unless the user accepts its agent usage. ### 7. Generate a deterministic fixture Repository generators may create functional example skills through the stable fixture contract: ```text node <skill-dir>/scripts/create-skill.mjs fixture \ --input "<fixture.json>" \ --repo-root "<target-repository>" \ --dry-run ``` Parse the JSON preview, obtain or propagate approval for its `hash`, then repeat the unchanged command with `--approve <hash>`. Fixture mode is always noninteractive. It accepts only the version 1 schema, deterministic built-in placeholder art, and the normal repository registration path. It never bypasses dry-run, approval, validation, atomic writes, or provenance. Read [fixture mode](references/fixture-mode.md) for the exact payload and exit contract. Repository generators use `create-skill register create-skill --dry-run`, followed by the unchanged command with `--approve <hash>`, after catalog staging. This registers the bundled authoring skill through the same atomic path without rewriting its skill files. ## Safety Invariants 1. Dry-run never writes files or calls an image provider. 2. Approval binds the exact inputs and target paths. Changed inputs require a new preview. 3. Provider requests make one attempt with no retry, redirect, or fallback. 4. Custom workflows require an exact preview and approval before any external action. 5. Provider output is untrusted until the complete PNG validator accepts it. 6. The validated skill thumbnail is the tracked source; catalog builds generate public copies. 7. Secrets never enter prompts, previews, logs, generated files, or provenance. 8. Existing files are preserved when planning fails. Apply uses staged sibling files and rollback. 9. No git staging, commit, push, issue, pull request, or repository setting change is part of this skill. ## Exit Criteria - New skills declare `metadata.version` as a quoted SemVer string. Preserve existing skill versions during registration; use the repository's version helper to bump a release and synchronize plugin listings when available. - The skill has every required file and only portable SKILL.md frontmatter keys. - Every discovered or configured registration surface includes the skill. - The tracked skill thumbnail passes strict PNG validation; catalog builds generate the public copy. - Prompt and non-secret provenance are recorded when that repository surface exists. - Deterministic tests and Vally lint pass. - The final report lists exact changed files, commands, results, and any blocked work. -
thumbnail.png 354.7 KB · in bundle
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.