Something went wrong. Try again.
A prompt to generate skills
Something went wrong. Try again.
6.0 kB · 118 lines · Created 8mo ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118copy and paste what's below here==================================================here's the structure for creating a claude skill. follow these rules and guidelines to a t.
Claude Skills: Power User GuideWhat a Skill Actually IsA skill is a folder with a SKILL.md at its root, plus optional bundled resources. Claude reads the SKILL.md when triggered and follows its instructions like a mini system prompt for that task. There are no variables, loops, or iterators in the traditional programming sense — the "logic" is natural language instructions telling Claude what to do conditionally, sequentially, or branching.The Core Structuremy-skill/├── SKILL.md ← required; triggers + instructions└── references/ ← optional deep docs loaded on demand└── assets/ ← optional templates, fonts, brand files└── scripts/ ← optional executable code Claude can runSKILL.md anatomy:markdown---name: my-skilldescription: [TRIGGER TEXT — this is what Claude pattern-matches against]---# Instructions...the actual directions...The Real "Variables" — The Description FieldThe description in the YAML frontmatter is the most important thing you'll write. Claude decides whether to load the skill based on this field alone. It's not code, but it functions like a matching rule.Weak description: "Helps with marketing copy"Strong description: "Write brand-aligned marketing copy, campaign briefs, launch announcements, ad copy, email sequences, or any customer-facing content. Use this skill whenever the user mentions campaigns, positioning, messaging, copy, launches, or asks to write anything that will be seen externally."Key tactics:Be pushy — skills undertrigger more than overtriggerInclude synonyms and phrasings users actually typeMention roles ("when a marketer asks..." or "for finance team...")Include both the what and when"Control Flow" in Natural LanguageSince there's no code interpreter in SKILL.md, you use structured prose as logic:Branching (if/else):markdown## Step 1: Identify Document Type- If the user says "memo" → load `references/memo-format.md`- If the user says "board update" → load `references/board-format.md`- If unclear → ask: "Is this internal or external?"Sequential (loops via numbered steps):markdown1. Draft the section2. Review against brand guidelines in `references/brand.md`3. Revise if any guideline is violated4. Repeat steps 1-3 for each sectionConditionals:markdownIf the output is longer than 1 page, add an executive summary at the top.```This is the mental model: **you're writing directions for a very capable intern, not a compiler.**---### Progressive Disclosure — The 3-Level SystemThis is the key architectural concept:| Level | What | Size | When loaded ||-------|------|------|-------------|| 1 | `name` + `description` | ~100 words | Always (every message) || 2 | SKILL.md body | <500 lines ideal | When skill triggers || 3 | `references/` files | Unlimited | Only when SKILL.md says to |**Practical implication:** Keep SKILL.md as a *router*, not an encyclopedia. Put deep content in reference files and tell Claude when to read them.---### Reference Files — How Many / How Long?- **SKILL.md body:** Under 500 lines. If you're going over, add a `references/` folder.- **Individual reference files:** 300+ lines → include a table of contents at the top so Claude can navigate efficiently.- **Number of files:** No hard limit, but organize by use case so Claude reads only what's relevant (e.g., `references/aws.md`, `references/gcp.md` — Claude reads only the one that matches the user's context).- **Assets (images, fonts, templates):** No size limit — Claude can use these without loading them into context.---### Startup Role Templates to BuildHere's how this maps across your team:**Product**```references/ prd-template.md user-story-format.md feature-spec.md```Trigger on: "write a PRD", "spec this feature", "user stories for..."**Marketing**```references/ brand-voice.md campaign-brief-template.md email-sequences.md launch-checklist.md```Trigger on: "campaign", "copy", "launch", "email sequence", "positioning"**Finance**```references/ financial-model-conventions.md board-update-format.md budget-narrative-template.md```Trigger on: "board deck", "budget", "financial summary", "runway analysis"**Ops/Recruiting**```references/ job-description-template.md offer-letter-format.md process-doc-template.mdPower User Patterns1. Chain skills via referencesSKILL.md says: "First read references/brand.md, then read references/channel-specific.md based on the platform."2. Scripts for deterministic workPut Python/bash in scripts/ for things like file transforms, data formatting, or API calls. Claude runs them without loading them into context (efficient).3. Asset injectionPut a Word/PPTX template in assets/ — the skill tells Claude to populate that template rather than build from scratch. Huge quality boost.4. Persona + constraints in SKILL.mdmarkdown## ToneWrite like a senior operator, not a consultant. Avoid jargon. Max 3 bullet points per section.Never use the word "leverage."5. Eval loop for iterationWrite 5–10 test prompts, run them, review outputs. The skill-creator skill has a full eval harness if you want to go deep.Quickstart for a Non-Eng SkillCreate my-skill/SKILL.md with a punchy description + 20–50 lines of instructionsAdd 1–2 references/ files with templates or style guidesTest by typing a prompt that should trigger it and see if Claude reads the skillIf it doesn't trigger — rewrite the description to be more specific and "pushy"
i want you to create a skill for (super verbose explanation of a particular job to be done. consider not only the artifacts to be directly generated, but also any related artifacts to prove quality or success)if you think you can also generate meaningful reference files, do so.