agentleFS
azborgonovoazborgonovo/ai-skills/.claude/skills/skill-authoring/SKILL.md

skill-authoring

Skill1 starsChanged 2 months agoMITView on GitHub

Conventions for writing or editing a SKILL.md in this repo — horizontal rules, bold usage, line wrapping, spelling, negative instructions, referencing other skills, when to reach for a scripts/ helper, and the bundled `check_skills.py` validator to run before committing. Use whenever creating a new skill or improving an existing one in this repository.

89 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use cloud-api-recipe-authoring, codebuddy-ide-mcp-upgrade and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • TencentCloudBaseSkill

    cloud-api-recipe-authoring

    Author or revise a cloud-api-operations recipe (config/source/skills/cloud-api-operations/references/recipes/). Encodes why the long tail of control-plane capabilities is deliberately NOT wrapped as MCP/CLI tools, the API-first admission test for whether a capability deserves a recipe, the negative and positive lists, the fixed recipe structure, the writing rules distilled from writing-for-agents, the user-facing copy red lines, and the product regeneration/verification pipeline. Also routes contributors (and their agents) to the contribution flow and the pre-PR self-test protocol in references/contributing.md. Use when adding a recipe, judging whether a scenario belongs in the recipe set, updating the recipes index, re-checking existing recipes against the admission test, or onboarding someone who wants to submit one.

    Also in 3 projects that use this

  • TencentCloudBaseSkill

    codebuddy-ide-mcp-upgrade

    升级 CodeBuddy IDE(genie 扩展)内置的 CloudBase MCP,以及 MCP 发版时同步 IDE 侧白名单。当用户提到「更新 IDE 里的 MCP」「内置 MCP 版本太老」「IDE 集成的 CloudBase 功能不足」「改工具白名单 toolWhiteList」「把新 bundle 打进 CodeBuddy」「白名单漂移」「MCP 发版要同步什么」时使用。覆盖:解包定位内置 bundle 与内嵌配置、重新构建 mcp bundle、生成新的工具白名单与系统提示词、安全注入 IDE 并备份、用 MCP 协议验证工具清单、人工端到端验收、一键回滚。

    Also in 3 projects that use this

  • DIL8654Skill

    api-contract-review

    Review ASP.NET Core Web API contracts for HTTP semantics, versioning, backward compatibility, and response consistency. Use when the user asks to review an API, check endpoints, validate REST behavior, or before releasing contract changes.

    Also in 5 projects that use this

  • zaxbysauceSkill

    pr-review-fix

    Codex adapter for addressing pull request review feedback in opencode-swarm. Use when the user asks to fix PR review comments, requested changes, reviewer findings, CI-review notes, or a pasted review summary with low false-positive tolerance.

    Also in 3 projects that use this

  • albertdobmeyerSkill

    git-workflows

    Advanced git operations beyond add/commit/push. Use when rebasing, bisecting bugs, using worktrees for parallel development, recovering with reflog, managing subtrees/submodules, resolving merge conflicts, cherry-picking across branches, or working with monorepos.

    Also in 3 projects that use this

  • rogue-dev-studioSkill

    frontend-engineering

    Expert frontend application engineering: accessible UI structure, state and data-fetching discipline, form/validation UX, responsive layouts, performance budgets, and integration with API contracts. Use when building or refining SPA/web screens, design-system consumption, client-side routing, or fixing UI reliability (loading, empty, and error states).

    Also in 3 projects that use this

  • +6 more from TencentCloudBase

The file

We split this skill into the rules it gives an agent

6 rules. 0 are common, 2 are distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 79 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Do not use --- as a section separator between steps or sections.

    line 10
  2. Very specific:

    Do not hard-wrap prose in a SKILL.md body.

    line 26
  3. Distinctive:

    Write the body in plain English, per Plain English in agent-facing markdown in the repository guidelines.

    line 32
  4. Distinctive:

    Prefer stating what to do over what to avoid.

    line 56
  5. Very specific:

    Prefer instructions that explain the *why* over a rigid wall of MUST and NEVER formatting rules, and keep the always-loaded SKILL.md body …

    line 113
  6. Very specific:

    When unsure, stay model-only and let evidence decide.

    line 117

“Common” is measured: kept by at least 6 in 10 of the 79 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

This repo holds Claude Code skills. Claude reads a SKILL.md file, and a human reads it only rarely. So formatting has to earn its tokens by helping the model parse and weight instructions, rather than by looking tidy. Apply these rules when you create a new skill or improve an existing one.

Horizontal rules (---)

Do not use --- as a section separator between steps or sections. The ## and ### Markdown headers already delimit sections.

The only --- that belong in a SKILL.md are the two YAML frontmatter delimiters at the very top. They open and close the name and description block.

Bold (**...**)

Bold is a signal, not decoration, and signal weakens when it sits everywhere. Use it on purpose:

  • Keep it for labeled lead-ins that act as inline sub-headers, such as **Truncation check**: and **Why the script**:. These help the model locate and weight a specific piece of guidance.
  • Keep it for directive anchors on things that change behavior, such as **Never** approve, **Always** post through the script, and **Critical**.
  • Drop it where it is purely cosmetic. Bold on a mid-sentence phrase that needs no emphasis dilutes the bold that does matter.

When in doubt, ask one question: does the model behave differently without this bold? When the answer is no, leave the text plain. Do not strip bold wholesale either. The loss of the useful lead-ins and directives costs more in clarity than it saves in tokens.

Line wrapping

Do not hard-wrap prose in a SKILL.md body. Write one line per paragraph, and one line per list item, and let the editor soft-wrap the display.

In the YAML frontmatter, a folded description: > block still wraps across indented lines.

Prose and spelling

Write the body in plain English, per Plain English in agent-facing markdown in the repository guidelines. That section carries the sentence limit, the tense and voice rules, and the two exemptions.

Use American English spelling, such as behavior, normalize, and canceled, so that terminology stays consistent across skills. The exception is text quoted word for word from another source. That covers an example string copied from a sibling skill, and the wording of a real product. Keep the original there rather than "correcting" the quote.

The description field

The description is retrieval text, and not prose. Claude reads it to decide whether to load the skill. State what the skill does, in the third person, and then state when to use it. That is the rule in Anthropic's skill authoring best practices, and every official example writes the second part as "Use when ...".

Keep that wording. Write the main trigger as Use when the user ..., and open a further branch with Use it also when .... Quote the phrasings that a real request carries. Do not replace the framing with a sentence about the triggering of the skill, as in "The trigger covers phrasings such as ...". That shape keeps the keywords and drops the instruction that Anthropic recommends around them.

A skill with disable-model-invocation: true is the one exception. Claude cannot load such a skill, so "Use when" promises something that cannot happen. Say that the skill is user-only, name its command, and then write the trigger as "When the user wants ..., suggest this command".

A plain English rewrite must not cost a description its framing. The exemption for a description in the repository guidelines exists for that reason.

Referencing other skills

Reference a sibling skill by name when it lives in the same plugin. Everything under skills/<plugin>/ installs as one unit, so that pointer always resolves for an installed user. Naming the sibling and handing off to it is cheaper and clearer than restating what it already covers.

Keep a skill in a different plugin out of the instructions. Plugins install independently, so the skill you point at can be absent, and the reader is left with a step that they cannot take. Where one genuinely helps, phrase it as an optional accelerator and keep the fallback local. Say what to do when it is absent. Define any vocabulary that the guidance needs here, rather than importing it from a plugin that nobody installs.

The description field is bound by the same rule, and not only where it states a boundary. A description that names the skill of another plugin ties the retrieval text of this skill to something that can be absent. That holds for a hand-off, for a scope boundary, and for a trigger phrase. State a boundary as what this skill does not do. Write a trigger from what the user says about their own situation, rather than from the tool that they happened to use to get there.

Counter-examples and negative instructions

Prefer stating what to do over what to avoid. A bare "do not do X" leaves X sitting in the context with no positive target to replace it. A clear positive rule with its reasoning is usually enough on its own.

Discard a skill that shows no measured lift

A skill earns its place only when its gains outweigh its cost. When an eval against the plain model shows no pass-rate gain, and the skill adds token or latency cost, discard it. Do not ship it, and do not keep it "just in case".

Precedent — consistency-first: a drafted skill for keeping generated code aligned with the existing patterns and vocabulary of a codebase. It ran across four eval iterations, on easy and hard fixtures, on Opus and Haiku, plus a drift scenario built as its best case. It produced essentially the same code as the plain model, and it spent roughly 10% more tokens. Current models already match neighboring code, read steering docs and ADRs unprompted, and re-anchor after local drift on their own. It was discarded on that evidence.

The thresholds that settle a verdict

Read the eval before you apply the rule above, because one criterion does not fit every skill. The tag on its cases sorts it. A behavior suite measures a reasoning skill against a no-plugin baseline arm. A mechanics suite measures a Manual skill, which a baseline arm cannot reach at all.

A reasoning skill earns its place through better thinking, so the delta over the baseline is the test. A mechanics skill earns its place by doing something a baseline cannot do, such as driving a tracker or filling a template. A delta measures nothing there, so the test is the absolute score on the with-plugin arm.

A grader marked arm: with-only is dropped from the baseline arm and excluded from the score in both arms, so the delta compares like for like. Mark a grader that way when a baseline can never satisfy it: a format it never saw, or a skill it does not hold. Under --ablation none nothing is excluded, so a mechanics suite scores those graders normally, which is what its verdict needs.

Read the Fired column before the delta. The with-plugin arm loads the plugin, and the model still decides whether to reach for the skill, so a case where the skill never fired scores a delta of zero for a reason that has nothing to do with the quality of the skill. decide measured this exactly: the two cases that fired returned +0.20 and +0.17, and the two that did not returned 0.00.

VerdictThreshold
Decommission a reasoning skillThe baseline score matches or beats the with-plugin score, on cases where the skill fired.
Decommission a mechanics skillThe score stays low after one fix attempt. A low baseline is never the reason.
Improve the skillThe delta is positive, and the with-plugin score sits below about 80%. One grader that fails across cases gives the same verdict, and it names the edit.
Improve the descriptionThe delta is near zero because the skill did not fire, or the triggering axis reports a low Fires or Holds count. The skill body is not the problem.
Promote Trial to AdoptA score of about 90%, with a positive delta.
Demote Adopt to TrialThe measured score does not support the claim.

Read a verdict off three runs, never one. The agent is not deterministic, so a single run reports luck as skill. scripts/run_evals.py runs each case three times by default, and scripts/eval_report.py reports two numbers from that: pass@k, the cases that passed at least once, and pass^k, the cases that passed every time. Read pass^k for a skill that must work every time, and pass@k for one where a single good answer is enough. A wide gap between the two names a flaky skill, and a flaky skill is not ready to promote whatever its mean score says.

A suite that reaches 100% stops telling you anything. When pass^k hits every case, the suite is saturated: add harder cases rather than reading the score as proof the skill is finished.

A pass rate of 0% is a broken case far more often than an incapable model. Before you act on one, read the transcript and check that the case is solvable, that the prompt names paths that exist in the sandbox, and that no grader demands something the granted tools cannot do.

scripts/run_evals.py runs every axis and writes the numbers, and scripts/eval_report.py turns them into the records that README.md reads. See Evals for the case layout, which lives under evals/<plugin>/<skill>/ at the repository root, and for the grader types.

Environment-specific values

Do not hardcode a value that varies across machines or users. That covers a filesystem path, a clone-root convention, a tool version, and a default port. A value that matches your own setup reads as generic, and it runs correctly only on a machine shaped like yours. Everyone else gets a silent wrong answer, or a confusing failure with no clue why.

Prefer discovering the value instead. Search a short list of plausible candidates, or derive it from something already known. Two examples are the remote URL of a repo and a config file already on disk. When discovery comes up empty, ask the user directly rather than guessing further. A question is cheap, and a wrong guess that "worked when I tested it" is not.

Precedent — clone-root assumption: a host-adapter skill for locating the local clone of a repo assumed that every user clones repos under ~/projects/<org>/<repo>. That was true for the machine it was written on. Someone copied the assumption word for word into a second skill while generalizing it for a different code host, and nobody questioned it either time. It then failed on the very repo it was written in, which lives under ~/Code/<org>/<repo> instead. The fix: probe a handful of common roots, then fall back to matching by remote URL, then ask.

Check the mechanical rules before committing

A machine settles several rules here, rather than a reading. Those rules are the description cap, the name charset, the two---- rule, the body length, and one line per paragraph. They drift in silence. Three descriptions in this repo sat over the 1024-character cap until a full review pass found them. Run the bundled validator over any SKILL.md that you touched, before the commit that lands it:

python .claude/skills/skill-authoring/scripts/check_skills.py

With no arguments it checks skills/ and .claude/skills/. Pass a path, either a tree or a single SKILL.md, to check a skill that you are authoring in another repo. It exits non-zero on any failure, and it names the file and the rule.

It is advisory: it reports rather than gates, and a clean run says only that the mechanical rules hold. Everything that it stays silent about is still yours to judge. That covers no-ops, duplication, whether the description reads as third person, and whether a British spelling is a quote that it must keep.

General principle

Prefer instructions that explain the why over a rigid wall of MUST and NEVER formatting rules, and keep the always-loaded SKILL.md body lean.

Reach for a bundled scripts/ helper, which loads only when it runs, when a task is genuinely deterministic and repeated every run. Weigh that honestly rather than defaulting to it. Much of the value of a skill is the model's judgment, which a rigid script cannot do. A script can also actively mislead, by anchoring the model to the wrong signal. An existing tool, such as Grep, ripgrep, or Glob, often already covers the deterministic part, with no code to maintain.

When unsure, stay model-only and let evidence decide. Once you watch several runs independently re-derive the same helper, that is the cue to extract it into scripts/.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.