AI Guides › Workbench
By Nigel Guy · 8 min read
You ask Claude for something, it hands back a polished, confident piece, and the tone of the reply says "done". So you treat it as done. The assumptions it made without asking stay hidden, and so do the facts it was unsure of. When you do ask for a change, it fixes that one sentence and leaves the paragraph that depended on it saying the opposite. It looks like finished work, but nobody has actually signed it off.
The rule: Claude's first pass is a proposal. It comes back labelled as a draft with its guesses listed, you decide, and only your decision makes it final.
The fix is a small kit: one skill that sets up the loop, plus three lighter ways to get the same behaviour when a skill is overkill.
| Tool | What it does | Cost / free tier at time of writing | Best for | The catch |
|---|---|---|---|---|
| Draft-check-finish skill | A reusable instruction file that Claude loads when your request matches its description | Skills work on Free, Pro, Max, Team and Enterprise. Anthropic's pricing page lists Pro in US dollars only, so check the £ price, including VAT, at checkout | Repeated work where you want the same review loop every time | It triggers on a description match, not on every message |
| Instructions for Claude | Account-wide standing instructions applied to all your chats | Included on every plan | A short always-on reminder | It applies everywhere, including quick questions you don't want drafted |
| Project instructions | Instructions that apply only inside one project | Free accounts get up to five projects | One client or one piece of work that needs sign-off | Only works inside that project |
| One-off prompt | You paste the loop at the start of a chat | Free | Trying the idea before you commit to it | Gone when the chat ends |
A skill is a folder containing a SKILL.md file. The file opens with a short header holding a name and a description, followed by the instructions themselves. Claude reads the description first and decides from it whether to load the rest, so the description does most of the work.
How to set it up
draft-check-finish. Save the text below inside it as SKILL.md. Anthropic's help centre writes the filename in lower case, but its own examples and the skill Claude builds for you both use SKILL.md.If you'd rather not touch files, you can open a chat, describe the loop and ask Claude to build the skill for you. It uses Anthropic's skill-creator to produce a skill file, which you then save.
What goes in the instructions. Before you upload, replace [REVIEWER] with your name or role, and [WORK_TYPES] with the kinds of work you want drafted, such as client emails, reports or Python scripts.
---
name: draft-check-finish
description: Returns substantial work as a labelled draft with flagged assumptions, waits for review, then revises the whole piece. Use for writing, research, analysis and code.
---
# Draft, check, finish
You are producing [WORK_TYPES] for [REVIEWER], who approves every deliverable.
Your first version is a proposal for them to judge. It is never the answer.
## Round 1: draft
1. Do the complete task to your best standard. Don't hand back an outline unless asked.
2. Make the first line "DRAFT 1 — awaiting your review".
3. Below the draft, add a "Review notes" block of no more than five bullets:
- Assumed: decisions you made without being told.
- Least sure of: facts, figures or choices you'd check first, and why.
- Would change the result: inputs you didn't have.
4. Stop. Ask [REVIEWER] to approve, amend or comment. Don't describe the work as complete.
## Round 2: finish
- Approved: return a clean version without the label or notes, headed "Final — approved".
- Feedback given: apply it, then reread the entire piece for knock-on effects
(cross-references, totals, tone, tests, conclusions that relied on the old wording).
List up to three other things you changed as a result. Label it "DRAFT 2" unless told to finalise.
- Feedback contradicts an earlier instruction: ask which one wins before revising.
## Rules
- Only [REVIEWER] can declare work final.
- Skip the draft step for quick factual questions, one-line replies, or when told "no draft".
- Never invent sources, figures or test results to make a draft look finished. Say what you did not check.
- For code, state which parts you ran and which you didn't.
## Before sending any draft, confirm
- The draft label is the first line.
- Every assumption appears in Review notes.
- Nothing in your reply says the task is done.
The description is 164 characters. The help centre caps descriptions at 200 characters for skills you upload to Claude. Anthropic's developer docs allow 1,024, but those limits are for the API, so stay under 200. Names should use lower-case letters, numbers and hyphens only, and must not include "claude" or "anthropic".
Check it worked. Start a new chat and ask for something that fits [WORK_TYPES]. Claude's thinking should show it loading the skill, and the reply should open with the draft label. If it doesn't, ask outright ("use my draft-check-finish skill") and then make the description more specific about when it applies.
Click your initials in the bottom left, choose Settings and find Instructions for Claude. Whatever you put there applies to every conversation. Keep it short and use it as a backstop that points to the skill:
For any substantial piece of [WORK_TYPES], return it labelled as a draft, list your assumptions and weakest points underneath, and wait for my verdict before calling it final. Answer quick questions directly. If I haven't given you something you need, ask; don't fill the gap with a guess.
Replace [WORK_TYPES] with the same list you used in the skill.
Use these when only one strand of your work needs sign-off, such as one client or one report. Paste the "Round 1" and "Round 2" sections from the skill into that project's instructions, and the loop then applies only to chats inside the project.
To try the loop before you build anything, paste this at the start of a chat:
You're helping me produce [DELIVERABLE] for [AUDIENCE]. We'll work in rounds.
1. Ask me for anything you need that I haven't given you. Don't guess.
2. Write a complete first version headed "DRAFT 1".
3. Underneath, list in up to five bullets: what you assumed, what you're least confident about, and what would change the result.
4. Wait for me. When I reply with changes, apply them, reread the whole thing for anything my change affects, and tell me what else you adjusted.
5. Only call it final when I say "approved".
Before sending each draft, check the label is there and every assumption is listed.
Replace [DELIVERABLE] with what you want, such as a supplier comparison, and [AUDIENCE] with who will read it.