--- name: creating-skills-ee945c65d158 disable-model-invocation: true description: Skill authoring and optimization. Use when creating new skills, editing existing skills, verifying skill quality, or optimizing trigger descriptions. --- # Creating Skills Skills are eval-driven process documentation. Build evaluations first, write the skill to pass them, then optimize trigger descriptions for activation accuracy. For comprehensive Anthropic guidelines, see [best-practices-reference.md](best-practices-reference.md). ## Eval-Driven Development ### Why Evals First A skill without evals is untested code. The eval proves the skill teaches the right thing and activates at the right time. ### The Claude A / Claude B Pattern Use two separate agent sessions to avoid self-confirmation bias: 1. **Claude A** (author) — writes the skill content 2. **Claude B** (tester) — runs evaluation scenarios cold, without seeing the authoring process This separation ensures the skill works for any agent, not just one primed by the authoring context. ### Minimum Evaluation Requirements Every skill needs at least **3 evaluation scenarios** before writing content: 1. **Baseline (no skill)** — Run scenarios without the skill. Document what the agent did wrong, what rationalizations it used (verbatim), which pressures triggered violations. 2. **With-skill pass** — Run the same scenarios with the skill loaded. The agent should now comply. 3. **Adversarial/edge case** — Test boundary conditions, near-misses, and rationalization attempts. Record concrete outputs, not impressions. Quote agent responses verbatim when documenting failures. ### Eval-Write-Eval Cycle ``` 1. Design 3+ realistic evaluation scenarios 2. Run baseline WITHOUT skill (Claude B) -> document failures 3. Write minimal skill addressing those specific failures (Claude A) 4. Run same scenarios WITH skill (Claude B) -> verify compliance 5. If new failure modes appear, add counters and re-test ``` Do not add content for hypothetical cases. Every section should trace back to an observed failure. ## Skill Types - **Technique** — Concrete method with steps (e.g., root-cause-tracing, condition-based-waiting) - **Pattern** — Way of thinking about problems (e.g., flatten-with-flags) - **Reference** — API docs, syntax guides, tool documentation ## Directory Structure ``` skills/ skill-name/ # letters, numbers, hyphens only SKILL.md # required — main instructions, under 500 lines reference.md # optional — detailed docs, loaded on demand examples.md # optional — usage examples, loaded on demand scripts/ helper.py # optional — executed via Bash, NOT loaded into context assets/ template.pptx # optional — output files (templates, icons, fonts), NOT loaded into context ``` **`assets/`** holds files used in skill output (brand logos, slide templates, boilerplate projects). Unlike references, assets are never read into context — they're copied, modified, or embedded in output. Keep references **one level deep** from SKILL.md. Nested references (SKILL.md → advanced.md → details.md) cause Claude to partially read files. For files over 100 lines, include a table of contents at the top. ## SKILL.md Frontmatter ```yaml --- name: skill-name-with-hyphens description: What it does. Use when [specific triggering conditions] --- ``` **Hard constraints:** - `name`: Max 64 chars. Lowercase letters, numbers, hyphens only. No "anthropic" or "claude". - `description`: Max 1024 chars. Non-empty. No XML tags. - Prefer gerund form for names: `processing-pdfs`, `analyzing-data`, `testing-code` **Optional frontmatter fields:** - `disable-model-invocation: true` — only user can invoke (not auto-triggered, saves context budget) - `user-invocable: false` — only Claude can invoke (background knowledge, hidden from `/` menu) - `allowed-tools` — restrict tools when skill is active (e.g., `Read, Grep, Glob`) - `context: fork` — run in isolated subagent context - `agent` — which subagent type when `context: fork` is set - `model` — override model for this skill (e.g., `haiku` for cheap tasks) - `effort` — `low`, `medium`, `high`, `max` (Opus 4.6 only) - `hooks` — lifecycle hooks scoped to this skill's execution - `paths` — glob patterns limiting when skill auto-activates (e.g., `["src/**/*.py"]`) - `shell` — `bash` (default) or `powershell` - `argument-hint` — shown during autocomplete (e.g., `[issue-number]`) **Description rules:** - **Third-person always.** Description is injected into the system prompt. - **CRITICAL: Description = when to use ONLY.** NEVER summarize the skill's process or workflow. Testing shows that when a description summarizes workflow, Claude follows the description instead of reading the skill body. - Include **both what the skill does AND when to use it**. - Include 5+ keywords from actual user workflows. - **Front-load the key use case** — descriptions are truncated at ~250 chars in skill listings. - **Be assertive about activation** — include phrases like "Make sure to use this skill whenever the user mentions X, even if they don't explicitly ask for it." **Context budget:** All skill descriptions share a pool of **~1% of context window (fallback ~8K chars)**. Every character in every description consumes from this shared pool. Skills with `disable-model-invocation: true` are NOT loaded into context at all. Override with `SLASH_COMMAND_TOOL_CHAR_BUDGET` env var. ## SKILL.md Body Structure ```markdown # Skill Name ## Overview What is this? Core principle in 1-2 sentences. ## Quick Reference Table or bullets for scanning common operations ## Core Pattern Before/after comparison (for techniques/patterns) ## Implementation Inline code for simple patterns, link to file for heavy reference ## Common Mistakes What goes wrong + fixes ``` **Default assumption: Claude is already very smart.** Only add context Claude doesn't already have. Challenge each piece: "Does Claude really need this explanation?" **Writing style:** Use imperative/infinitive form ("Validate the input", "Configure the server"), not second person ("You should validate..."). Explain *why* behind instructions — LLMs respond better to reasoning than bare ALWAYS/NEVER directives. ## Degrees of Freedom Match specificity to task fragility: | Level | When to use | Example | |-------|------------|---------| | **High** (text instructions) | Multiple valid approaches, context-dependent | Code review guidelines | | **Medium** (parameterized scripts) | Preferred pattern exists, some variation ok | Report generation with template | | **Low** (exact scripts, no params) | Fragile operations, consistency critical | Database migrations, deployments | ## Token Efficiency - **Every-session skills** (loaded via hook): absolute minimum tokens. <200 words. - **Frequently-triggered skills**: <500 words preferred. - **On-demand skills** (`disable-model-invocation`): more generous, but still under 500 lines. - **Body word count target**: 1,500-2,000 words ideal, <5,000 words max. - Scripts in `scripts/` are executed, not loaded — only output consumes tokens. ### Progressive Disclosure Split Even skills under 300 lines benefit from splitting SKILL.md into a lean core + reference.md. The split reduces context consumption by ~46% on average with zero quality loss — Claude reads reference.md on demand only when it needs detailed code examples. **What stays in SKILL.md** (always loaded when skill triggers): - Workflow steps, decision tables, rules, API summaries - Minimal code snippets (2-5 lines) showing correct/wrong patterns - Checklists, quick-reference tables - Cross-references to other skills - A "## Detailed Reference" section linking to `[reference.md](reference.md)` **What moves to reference.md** (loaded only when Claude needs implementation details): - Full code examples (>10 lines) from the codebase - Complete class/system implementations showing patterns end-to-end - Deep-dive explanations of internal mechanisms - Test code examples **When to split:** Always consider splitting if the skill has code examples >10 lines. The question is not "is this skill too long?" but "does Claude need this code block to understand the workflow, or only when implementing?" If the latter, extract it. ## Description Optimization The description controls activation rate. Optimize methodically: ### Step 1 — Keyword Coverage Use words Claude would search for: error messages, symptoms ("flaky", "hanging", "race condition"), tool names, synonyms. Cover the vocabulary space of likely user prompts. ### Step 2 — Build Eval Queries Create at least 20 evaluation queries: - 8-10 should-trigger queries (realistic user prompts with personal context, specific details) - 8-10 should-not-trigger queries (near-misses from adjacent domains, not obviously irrelevant) ### Step 3 — Test and Iterate 1. For each query, check if the description would cause activation 2. Adjust keywords and phrasing to improve precision and recall 3. Re-test until satisfied with activation accuracy ### Activation Rate Benchmarks - **Good:** >80% correct activation on eval set - **Acceptable:** >70% with no false positives on critical should-not-trigger queries - **Needs work:** <70% or any false positive on dangerous near-misses ### Step 4 — Add Concrete Examples If keyword coverage alone is insufficient, add specific phrases from real user prompts. Concrete examples improve activation from ~50% to 72-90%. ## Workflow Patterns ### Checklist Pattern For complex multi-step tasks, provide a copyable checklist Claude can track progress with. ### Feedback Loop Pattern Run validator → fix errors → repeat. Greatly improves output quality for document generation, form filling, code review. ### Plan-Validate-Execute Pattern Have Claude create an intermediate plan file → validate with script → execute. Catches errors early. ### Conditional Workflow Guide Claude through decision points: "Creating new? → Creation workflow. Editing existing? → Editing workflow." ### Template Pattern Provide output format templates (strict or flexible) so Claude produces consistent structured output. ### Examples Pattern Input/output pairs work like few-shot prompting — show what good output looks like for representative inputs. ## Bulletproofing Discipline Skills For skills that enforce rules, resistance to rationalization is critical. **Close every loophole explicitly** — state the rule, then forbid specific workarounds. **Build a rationalization table** from baseline testing: ```markdown | Excuse | Reality | |--------|---------| | "Too simple to test" | Simple code breaks. Test takes 30 seconds. | ``` **Red flags list** — makes agent self-checking easy. ## Dynamic Features ### String Substitutions | Variable | Description | |----------|------------| | `$ARGUMENTS` / `$N` | Arguments passed when invoking the skill | | `${CLAUDE_SKILL_DIR}` | Directory containing the skill's SKILL.md | | `${CLAUDE_SESSION_ID}` | Current session ID | ### Dynamic Context Injection The `` !`command` `` syntax runs shell commands before skill content is sent to Claude: ```yaml ## PR context - PR diff: !`gh pr diff` - Changed files: !`gh pr diff --name-only` ``` ## Anti-Patterns - **Narrative** — "In session 2025-10-03, we found..." is too specific, not reusable - **Multi-language dilution** — One great example beats mediocre examples in 5 languages - **Generic labels** — `helper1`, `step3` have no semantic meaning - **Skipping evals** — Writing content before observing failures produces bloated skills - **Graphviz diagrams** — Rendered as raw text in terminals. Use numbered lists or tables instead. - **Workflow summary in description** — Claude follows description shortcut instead of reading body - **Vague descriptions** — "Helps with documents" achieves ~20% activation - **Over-explaining** — Claude knows what PDFs are and how libraries work - **Deeply nested references** — Keep one level deep from SKILL.md - **Time-sensitive info** — "Before August 2025, use old API" becomes wrong. Use "Old patterns" section. - **Inconsistent terminology** — Pick one term ("endpoint" not "endpoint/URL/route/path") - **Too many options** — Provide one default with an escape hatch, not 5 choices - **Excessive ALWAYS/NEVER caps** — Explain reasoning instead; LLMs respond better to "why" than bare directives ## Testing Across Models What works for Opus may need more detail for Haiku. If a skill will be used across models: - **Haiku**: May need more explicit guidance and examples - **Sonnet**: Balanced — good default target - **Opus**: May be over-served by verbose instructions — trim for efficiency ## Checklist **EVAL Phase:** - [ ] Design 3+ realistic evaluation scenarios - [ ] Run WITHOUT skill (Claude B) — document exact failures verbatim - [ ] Record rationalizations and failure modes **WRITE Phase:** - [ ] Name: lowercase, hyphens, max 64 chars, gerund form preferred - [ ] Frontmatter max 1024 chars, description in third person - [ ] Description: no workflow summary, includes "Use when..." + keywords, under 250 chars - [ ] Body under 500 lines. Heavy reference in separate files. - [ ] Appropriate degree of freedom for the task type - [ ] Only context Claude doesn't already have - [ ] References one level deep, files >100 lines have TOC - [ ] Progressive disclosure split — code examples >10 lines moved to reference.md **VERIFY Phase:** - [ ] Run WITH skill (Claude B) — verify compliance on all scenarios - [ ] Add rationalization table (if discipline skill) - [ ] Re-test until bulletproof **OPTIMIZE Phase:** - [ ] Build 20+ eval queries for description - [ ] Test activation accuracy (target >80%) - [ ] Add concrete examples if activation <70% - [ ] Test with Haiku, Sonnet, and Opus if cross-model use planned --- Source attribution: decentraland Original project: https://github.com/decentraland/unity-explorer/blob/9a03ce0931a819ca98557660d4d7531009ce100c/.claude/skills/creating-skills/SKILL.md License: Apache-2.0 License information: https://www.apache.org/licenses/LICENSE-2.0 Original upstream instructions; downloaded from Dave AI. ## Upstream license notice Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 1. Definitions. "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License. "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work. 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. END OF TERMS AND CONDITIONS APPENDIX: How to apply the Apache License to your work. To apply the Apache License to your work, attach the following boilerplate notice, with the fields enclosed by brackets "[]" replaced with your own identifying information. (Don't include the brackets!) The text should be enclosed in the appropriate comment syntax for the file format. We also recommend that a file or class name and description of purpose be included on the same "printed page" as the copyright notice for easier identification within third-party archives. Copyright [yyyy] [name of copyright owner] Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. # Dave AI resource metadata Creator / publisher: decentraland Original project: https://github.com/decentraland/unity-explorer/blob/9a03ce0931a819ca98557660d4d7531009ce100c/.claude/skills/creating-skills/SKILL.md Attribution basis: Source repository record Fields of work: Content & Media, Product Management Roles: Growth Product Manager, Product Manager, Technical Writer Specialties: Specialty not specified Tools: Tool not specified Country applicability: Country not specified Classification basis: Source catalog labels Instruction SHA-256: db87c5d51f54625179a765df84c1b4b7e4f207380725c77f974f145e1b042462