AGENT SKILL

skill-creator

Design a new skill or revise an existing skill's workflow, trigger boundaries, resources, or packaging. Use writing-for-agents for wording-only changes and skill-improvement for behavioral evaluation.

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
skill-creator
license
Complete terms in LICENSE.txt
description
Design a new skill or revise an existing skill's workflow, trigger boundaries, resources, or packaging. Use writing-for-agents for wording-only changes and skill-improvement for behavioral evaluation.

02 资源文件 8

03 SKILL.md 正文

Skill creator

Build a skill around a repeatable task that the agent needs help performing. Keep the user's scope and existing authorization intact.

Inspect and choose the scope

  1. Read the closest existing skill, its callers and bundled resources. Reuse or extend it when it already owns the capability.
  2. Define the requests the skill should handle and any likely neighboring task it should leave alone. Ask only for information that cannot be inferred from the request or existing artifacts.
  3. Before editing, fix a representative task, an edge case and an unused hold-out. Include at least one critical correctness requirement. For trigger changes, include a positive and a negative prompt.
  4. Check ownership. In Harness, edit the harness SSOT side. Rulesync-managed upstream skills remain unchanged; their updates come from rulesync.

Structure the skill

Keep purpose, shared steps, essential constraints, completion criteria and resource-loading conditions in SKILL.md. Put substantial mode-specific procedures, examples and schemas in references/, executable helpers in scripts/, and output resources in assets/. A short self-contained skill needs no extra directories.

Use relative links from SKILL.md and state when each file is needed. Read only relevant references; moving content out and then loading every file does not reduce context use. Keep each instruction in one authoritative place. Link to existing policies instead of copying their full checklists.

Treat frontmatter metadata as routing metadata, not as a miniature prompt. The description should identify the task boundary and distinguish neighboring skills; workflow steps and generic operating guidance belong in the body or conditional references. Do not require callers to preload architecture docs, examples, schemas or sibling skills unless every invocation needs them. Prefer explicit conditions such as “read X when Y is true”. Express completion as observable outcomes and verification rather than prescribing unnecessary intermediate behavior. Runtime/model-specific scaffolding belongs in target-specific policy only after an observed behavioral difference justifies it.

  • Trigger boundaries, context loading or metadata routing: read routing-and-context.
  • Multiple modes or complex resource layout: read resource-patterns.
  • Sequential or conditional workflow design: read workflows.
  • Output format or examples: read output-patterns.
  • Wording, pointers or hierarchy: use writing-for-agents; where it is not installed, review the same points yourself.

Aim below 500 lines and about 5,000 tokens for the entrypoint; these are upper guidelines, not targets. Metadata loads during discovery, so keep the description short and discriminating. Bytes and words are not token counts.

Write the workflow

Include task-specific guidance that changes decisions: fragile operations, known boundary conditions, required inputs and how to verify the result. Use a fixed sequence when order affects correctness. For flexible work, describe the outcome and selection criteria. Avoid generic tutorials, forced interviews, repeated safety text and speculative edge cases. Preserve supported metadata and existing invocation policy when updating a skill.

Frontmatter needs a nonempty name and description. Standard optional fields include license, compatibility, metadata and experimental allowed-tools. Claude also supports fields such as disable-model-invocation, user-invocable, argument-hint, context, agent, model and hooks. Runtime-specific fields do not imply the same behavior on other hosts. In Harness, inspect the target's frontmatter transform and per-target distribution declarations before claiming cross-runtime support. Keep tool-specific requirements in compatible targets or express the necessary outcome in portable prose.

For a new standalone skill, use the bundled initializer when appropriate:

python3 scripts/init_skill.py <skill-name> --path <destination>

Before invoking the initializer, require a single skill name matching ^[a-z0-9]+(?:-[a-z0-9]+)*$; reject separators, .., absolute paths and shell metacharacters. Confirm the resolved destination remains a child of the chosen root. The bundled initializer does not enforce this boundary itself.

Resolve the script path from this skill directory. Inspect the generated files, keep resources that the task needs, and remove scaffolding with the user's existing authorization. For an existing skill, edit its files directly rather than rerunning the initializer. For a repository with its own layout and distribution scripts, copy the closest local implementation instead.

Validate and evaluate

Run changed executable helpers against realistic inputs. Use the project's tests and distribution checks when available. For standalone frontmatter validation:

python3 scripts/quick_validate.py <skill-directory>

This is a basic syntax/name/field check; it does not validate every reference or prove behavioral quality. Open the referenced resources to verify paths and loading conditions. Confirm that mandatory inputs, outcomes and stopping conditions are reachable from the entrypoint.

Use skill-improvement for the baseline, fresh scenario runs and hold-out when workflow, triggers, reference routing or agent decisions change. Pass raw task inputs to evaluators and keep expected answers out of their prompts. Report failed or unavailable evaluations separately from completed checks.

Deliver

Use the repository's distribution mechanism for repository-managed skills. Harness distributes skill directories through target manifests; a .skill archive is unnecessary for that workflow. Only when the user needs an archive, run:

python3 scripts/package_skill.py <skill-directory> [output-directory]

Before packaging, reject symlinks in the skill resources and choose an output directory outside the skill root so the archive cannot include itself. The packager validates frontmatter and writes a ZIP-compatible .skill file. Inspect the archive before sharing it; packaging does not verify the workflow or grant publication approval.

Done means the skill routes the intended requests without stealing its nearest negative case, only loads resources needed for the active path, exposes a reachable completion boundary, passes available repository validation, and reports any unverified behavior. Report the changed behavior, validation and remaining limitations. Retain license and attribution files.

References: Agent Skills specification, skill creation best practices.

04 同仓库更多 Skills

harunonsystem
harunonsystem/harunon-harness-slim

writing-for-agents

Writing documents for agents. Use when creating or editing skills, or modifying AGENTS.md or CLAUDE.md.

★ 0
harunonsystem
harunonsystem/harunon-harness-slim

wrangler

Run or troubleshoot Wrangler CLI commands in projects that have a wrangler.jsonc, wrangler.json, or wrangler.toml file, including local development, Previews, deployment, and resource management. For new projects and projects without a Wrangler configuration file, use the cf CLI instead.

★ 0
harunonsystem
harunonsystem/harunon-harness-slim

workflow-modernize

パッケージ/CLI更新時や定期見直し時に、自作スキル・ワークフロー・設定を監査し、不要なworkaroundの撤廃・新機能の適用・プロンプトの簡素化をGrill(壁打ち)と多角比較で提案・修正する。「workflow-modernize」「スキル監査」「ワークフロー改善」「パッケージ更新後の見直し」「モダナイズ」で起動。

★ 0
harunonsystem
harunonsystem/harunon-harness-slim

workers-best-practices

Cloudflare Workers best practices for production applications. Use when writing, reviewing, or configuring Workers.

★ 0
harunonsystem
harunonsystem/harunon-harness-slim

web-perf

Audit, diagnose, or optimize website loading and interaction performance, Core Web Vitals, and Lighthouse performance scores.

★ 0
harunonsystem
harunonsystem/harunon-harness-slim

wayfinder

Plan a huge chunk of work (more than one agent session can hold) as a shared map of decision tickets on your issue tracker, and resolve them one at a time until the way to the destination is clear.

★ 0