AI Guides › Workbench
By Nigel Guy · 8 min read
You wrote the prompt, read it back, and it made perfect sense. That is the problem: you read it with everything in your head that Claude does not have, so every gap filled itself silently. The usual response is to add more adjectives or paste in a clever template, when the fix is much cheaper: find out whether someone with none of your context could do the job from the page alone.
The rule: before you send a prompt that matters, hand it to a stranger, human or model, and fix every place they stop or guess.
The idea is not ours. Anthropic's prompting guidance describes the same habit: imagine handing your draft to a colleague who knows little about the job, and ask yourself whether they could act on it. If they would be puzzled, expect Claude to be puzzled as well. The guidance also suggests picturing Claude as a very capable new starter who has not yet learned how your team works. Below are four practical ways to put that habit to work.
Claude cannot see your inbox, your last three meetings, your house style or what "the usual format" means to you. It only has the words you sent (plus any files, project instructions or memory your setup supplies, which you should not assume are doing what you think). A new colleague is in the same position. Every undefined term, unstated audience and missing "what good looks like" is a guess waiting to happen, and a capable reader guesses fluently, so the output looks confident even when it solved the wrong problem.
The stranger test exposes those guesses before you spend a run on them. It is also quick: reading a prompt cold takes a minute, and rewriting it takes a few more.
| Tool | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| 1. The human cold read | A real person follows your prompt with no briefing | Free (a few minutes of their time) | Prompts you will reuse or hand to a team | Needs someone willing; they may be too polite to say "I'd guess" |
| 2. The cold-read prompt | A fresh Claude chat plays the stranger and lists what it would have to guess | Free beyond your normal plan usage; check your own plan's limits | Quick checks before any important run | A model fills gaps well, so ask it to list them rather than quietly cope |
| 3. The gap-fixing rewrite | Turns the gap list into a tighter prompt, asking you for missing facts | Same as above | Prompts you keep reusing | It can only fix gaps you answer honestly |
| 4. The restate-first run | Makes Claude say what it understands before doing the work | Same as above | One-off tasks where a wrong turn is expensive | Adds one round trip |
No tool here has a subscription price of its own; they run inside whatever Claude plan you already use.
Pick someone who was not in the room: a colleague from another team, a friend, a family member for a non-technical task. Give them only the prompt text and any files you would attach. Ask them to do the task, or to say out loud what they would produce, and to stop and tell you the moment they would be guessing. Do not explain anything. Every time you open your mouth to clarify, that clarification belongs in the prompt.
Open a new chat with no project, memory or earlier messages, so the reader really is a stranger. Paste this, then paste your prompt where indicated.
You are a capable new colleague who has just joined with no background on my work. I will give you a prompt I plan to send to an AI assistant. Do NOT carry out the task in it.
Your job is to read it cold and report what is unclear.
Context: the prompt is for [WHAT_THE_TASK_IS_FOR] and the AI will see only the prompt text and [ATTACHED_FILES_OR_NONE].
Work in this order:
1. Say in two sentences what you think the task is and who the result is for.
2. List every term, name, abbreviation or reference you could not be sure about.
3. List every decision you would have to guess: audience, length, tone, format, what to include or leave out, what "good" means.
4. List any instructions that clash with each other.
5. Name the one missing fact that would most change the result.
Rules: do not fix the prompt yet, do not invent answers to your own questions, and if something is clear, say nothing about it. Before replying, check that each point you raise is a genuine gap and not a matter of taste.
Here is the prompt:
[PASTE_YOUR_PROMPT]
Fill in: what the task is for, what the assistant will actually have access to, and your prompt.
Take the list from Tool 2 back to the original chat, or start another, and use this.
You are an experienced prompt editor. Below are my draft prompt and a list of gaps a cold reader found. Rewrite the prompt so a newcomer could follow it without asking questions.
Draft prompt:
[YOUR_DRAFT_PROMPT]
Gaps found:
[GAP_LIST_FROM_COLD_READ]
Facts I can confirm now: [ANSWERS_YOU_ALREADY_KNOW]
Steps:
1. For any gap that needs a fact only I have (audience, deadline, preferred format, examples of good work), ask me first, in one numbered list. Do not guess them.
2. Once I answer, rewrite the prompt with: a one-line role, the purpose and audience, the numbered steps in order, the output format, constraints, and what to do if information is missing.
3. Put anything variable in [SQUARE_BRACKET_PLACEHOLDERS].
4. Keep my own wording where it was already clear.
Check before replying: could someone with no context follow this? Is every term defined? Have you added any requirement I did not ask for? Remove it if so.
Fill in: your draft, the gap list, and any answers you already have. This also reflects Anthropic's advice to explain why an instruction exists instead of issuing a bare rule; its illustration is noting that the reply will be spoken by a text-to-speech system, which is why ellipses cause trouble.
For a task where a wrong turn costs real time, add this to the end of your prompt instead of running a separate check.
Before you start, reply only with: (a) the task in your own words, (b) the three assumptions you are making that I have not confirmed, and (c) any question you need answered. Then wait for my go-ahead.
Task: [YOUR_TASK]
Fill in your task. If the restatement is off, you have found your gap for the price of one message.
Whichever tool you use, read the result against this card.
| Check | A stranger stalls when... |
|---|---|
| Goal | They cannot say what the result is for or who reads it |
| Terms | Your shorthand, project names or acronyms are undefined |
| Steps | The sequence matters but you never set it out; numbered steps are the usual remedy, and Anthropic's guidance leans the same way |
| Format | Length, structure and tone are left to taste |
| Examples | You describe a style rather than showing one; a handful of varied, relevant samples in tags, as Anthropic advises, settles format and tone |
| Missing input | The prompt refers to a document, number or decision it never supplies |
| Success | There is no way to tell a good answer from a plausible one |
The trap is passing the test with a reader who is not a stranger. A chat that already holds your project, memory or earlier messages has the context you are trying to check for, so it will sail through. Use a clean chat, and remember that your real run may have different context (project instructions, files) than the test, so test the prompt in the setting it will actually run in too.
Run the test when the prompt will be reused, when it is long, when it mixes instructions with pasted material, or when a previous answer came back competent but wrong.