AI Guides › Workbench
By Nigel Guy · 5 min read
You type the task, get a flat answer, and then spend six messages steering it towards what you meant all along. The usual diagnosis is "the AI isn't good enough". The more common cause is that the one sentence carrying your real intent never left your head. It felt obvious to you, so it didn't get typed.
The rule: if a new colleague with no context couldn't do your task from the prompt alone, add the one sentence they would have asked about.
That rule is not ours. Anthropic's current prompting guide frames Claude as "a brilliant but new employee who lacks context on your norms and workflows", and its "golden rule" is to show your prompt to a colleague with minimal context and see whether they'd be confused. It also says that giving the reason behind an instruction helps the model deliver more targeted responses. This guide turns that into five reusable sentences.
A note on scope: we couldn't identify one specific "magic sentence" behind the social-media posts that circulate under this idea, so we haven't tried to defend one. What follows is the verifiable, durable part: which kinds of missing sentence matter, and where to put them.
| Sentence type | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| Purpose | Says what the answer will be used for | Free | Anything you'll send, publish or decide on | Vague purposes ("for work") add nothing |
| Reader | Names who will read it and what they already know | Free | Emails, explainers, summaries | Don't invent a reader you don't have |
| Finish line | Says what "done" looks like and the format | Free | Reports, lists, comparisons | Over-specifying length can crowd out substance |
| Ask first | Tells the model to question you before answering | Free | Big, ambiguous tasks | Wastes a turn on small tasks |
| Self-check | Asks for a check against your criteria before it replies | Free | Facts, maths, code, anything with a rubric | Doesn't catch errors it can't see; check sources yourself |
All five are free to use; they cost a few seconds of typing and, in some tools, a few extra tokens of output. Check your own plan if you pay per token.
The Anthropic guide's own example: "NEVER use ellipses" works less well than explaining that the text will be read aloud by a text-to-speech engine that can't pronounce them. The reason lets the model generalise to cases you didn't list.
Write [THING_YOU_NEED]. I'm going to use it to [WHAT_YOU_WILL_DO_WITH_IT], so [WHY_THAT_CHANGES_THE_ANSWER].
Fill in the thing, the use, and the consequence (for example "so it must stand alone without the attachment").
You are helping me write for [WHO_WILL_READ_IT]. They already know [WHAT_THEY_KNOW] and do not know [WHAT_THEY_DONT]. Write it so they can [ACTION_THEY_SHOULD_TAKE] after one read.
Anthropic advises being specific about the output format and constraints, and using numbered steps when order matters.
Task: [TASK].
A good result is: [SUCCESS_DESCRIPTION].
Format: [FORMAT, e.g. a table with columns X, Y, Z / five bullets / 200 words].
Leave out: [WHAT_YOU_DON'T_WANT].
For anything big, make the model gather what it's missing instead of guessing.
Role: you are an experienced [ROLE] helping me with [TASK].
Context: [CONTEXT].
Before you answer, list the facts you still need from me, as short numbered questions (maximum [NUMBER]). Do not draft anything until I have replied. If nothing is missing, say so and proceed.
Anthropic's guide recommends asking the model to verify its answer against stated criteria before finishing, noting it helps especially with coding and maths. It also says one current model, Claude Opus 5, verifies well without being told, so on that model such instructions can add cost and delay. Treat this one as optional and model-dependent.
Before you give your final answer, check it against these criteria: [CRITERIA]. If it fails any, fix it first. Then tell me in one line which checks you ran and anything you could not verify.
| Where | Use it for | Notes |
|---|---|---|
| In the message | One-off tasks | The default. Put inputs in clear labelled blocks. |
| Claude: Settings, then "Instructions for Claude" | Sentences you want on every chat, such as your reader or format habits | Per Anthropic's support page, available on all plans at time of writing. |
| Claude: project instructions | Context for one body of work | Applies only to chats in that project. |
| ChatGPT or other tools | Same idea: a saved-instructions or system-level slot | Menu names change; check the tool's current help page rather than trusting a screenshot. |
Standing instructions suit the Reader and Finish line sentences. Keep Purpose and Ask first in the message, because they change from task to task.
Add one sentence at a time. If you bolt on all five to every message you'll get a bloated prompt and won't know which sentence did the work.