AI Guides › Playbooks
By Nigel Guy · 7 min read
The usual way to build a "creative studio" in Claude is to write one enormous instruction file that covers research, ideas, drafting, editing and formatting, then wonder why the output is generic. It feels thorough while you write it. In practice Claude gets a wall of competing instructions and averages them. Paying someone thousands to wire up a bigger system rarely fixes that, because the problem is the shape, not the size.
The rule: one skill per stage of the work, each with one job, one input and one output that the next stage can pick up. If a skill needs the word "and" in its description, split it.
A note on sources. This topic arrived framed as a specific "seven skills" studio. I could not verify that framing against any official source, so I have not reproduced it. What I can verify is how Claude Skills work, and the seven stages below are my own design on top of that. Treat the stage list as a starting layout, not a standard.
A skill is a folder containing a SKILL.md file with a name and a description, plus optional extra files and scripts. According to Anthropic's documentation, Claude loads only the name and description at the start (roughly 100 tokens per skill). The full instructions load only when your request matches the description, and bundled files load only when they are read. So many small skills cost little, while one giant skill is loaded in full every time it triggers.
The description is the trigger. It must say what the skill does and when to use it. Vague descriptions are the most common reason a skill never fires.
Seven stages, one skill each. Each row names what goes in and what must come out.
| # | Skill | Input | Output (hand-off) |
|---|---|---|---|
| 1 | brief-intake | A rough request | A one-page brief: audience, goal, constraints, success test |
| 2 | research-pass | The brief | Sourced notes, with unverified items marked |
| 3 | idea-spread | Brief and notes | Five distinct angles, each in one sentence, with one recommended |
| 4 | first-draft | The chosen angle | A full draft in your house format |
| 5 | critique-pass | The draft and the brief | A numbered list of faults against the brief, no rewriting |
| 6 | channel-adapt | The approved draft | Versions for each channel you actually use |
| 7 | final-check | Everything | A pass/fail checklist: facts, claims, tone, legal, links |
The ledger is the mechanism. Each skill reads the previous output, never the whole history, which is what keeps stages from bleeding into each other.
Start with stages 1, 5 and 7: the brief, the critique and the final check. They are the cheapest to build and the ones most people skip. A draft produced against a written success test, then criticised against that same test, improves more than any amount of prompt polish on the drafting step. Add research, ideas, drafting and adaptation one at a time, and only when you notice that stage failing.
On a Free, Pro or Max plan, Anthropic's help centre says to turn on "Code execution and file creation" under Settings > Capabilities, then go to Customize > Skills, choose the plus button, then "Create skill", then "Upload a skill", and upload a ZIP. The ZIP must contain the skill folder as its root, not loose files. Team and Enterprise owners enable skills in Organization settings. Menu names and plan availability differ slightly between Anthropic's own pages (one developer page says Settings > Features and lists Pro, Max, Team and Enterprise), so check what your own account shows.
Naming rules from the developer documentation: the name uses lowercase letters, numbers and hyphens only, up to 64 characters, and cannot contain "anthropic" or "claude". The description is capped at 1,024 characters there, although an older help article says 200. Keep it under 200 and you satisfy both.
Cost at time of writing: Pro is listed at $20 a month, or $17 on annual billing (about £15 at time of writing; check the £ price at checkout). Max starts from $100 a month (about £75 at time of writing). Claude Code users can instead keep skills as files in ~/.claude/skills/.
Use this prompt to have Claude draft a skill for you, then edit it yourself.
You are helping me write one Claude skill (a SKILL.md file) for a single stage of a creative workflow.
Stage: [STAGE_NAME, e.g. critique-pass]
What it receives: [INPUT]
What it must produce: [OUTPUT_FORMAT]
My house style and constraints: [STYLE_AND_RULES]
One example of a good result: [EXAMPLE_OUTPUT]
Things it must never do: [DO_NOT_LIST]
Write the SKILL.md in this order:
1. Frontmatter: a lowercase-hyphen name, and a description under 200 characters that says what the skill does and when to use it.
2. A short purpose line.
3. Numbered steps.
4. The exact output format, with a template.
5. Guardrails, including: ask me for any missing input instead of guessing; mark any claim you cannot support as unverified.
Keep it to one job. If I have described two jobs, tell me and propose a split instead of writing it.
Before you answer, check the description would not also trigger on neighbouring stages, and say if it might.
Fill in the bracketed fields. Do the same for each stage.
A hypothetical scenario: a small bakery wants a launch announcement for a new sourdough subscription.
Notice that stage 5 never rewrites. Critique and drafting are separate skills because a model that critiques and fixes in one go tends to fix the easy things and defend the draft.