AI Guides › Workbench
By Nigel Guy · 6 min read
Most prompts grow a tail of prohibitions. "Don't be wordy. Don't use bullet points. Don't make things up. Don't sound like a robot." It feels like control, because each line closes a door you have already been burned by. The trouble is that a prohibition tells Claude only what to avoid, so it has to guess what you want instead, and the guess is often something you did not want either.
The rule: for every "don't" in a prompt, write down the thing you want in its place, and keep the "don't" only if it guards a real line and carries a reason.
Anthropic's current prompting guide, in the section on controlling response format, lists "Tell Claude what to do instead of what not to do" as the first technique. Its own example: instead of "Do not use markdown in your response", try "Your response should be composed of smoothly flowing prose paragraphs."
The same guide has a second, related point. A bare command such as "NEVER use ellipses" is described as less effective than giving the reason: the response will be read aloud by a text-to-speech engine, which cannot pronounce ellipses. The guide adds that Claude is smart enough to generalise from the explanation. It also advises dialling back shouting: where you once wrote "CRITICAL: You MUST use this tool when...", plain wording such as "Use this tool when..." now works, because newer models respond more strongly to the system prompt.
What the guide does not say is that negatives are banned. Its own example for stopping preamble reads "Do not start with phrases like 'Here is...', 'Based on...'", which is a prohibition with specific wording. I could not find an official explanation of why a prohibition performs worse at a mechanical level, so treat "backfire" as a practical tendency, not a proven law. The reliable part is the direction of the advice.
| Tool | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Swap | Turns a "don't" into the behaviour you want | Free | Style, tone, format, length | You must know what you want |
| Reason clause | Adds "because..." so Claude can handle cases you did not list | Free | Rules with a real purpose | Needs an honest reason |
| Kept line | Keeps a prohibition, made specific | Free | Hard limits, safety, accuracy | Overuse brings the problem back |
| Rewriter prompt | Has Claude propose swaps for your existing prompt | Free to write; uses your normal Claude usage | Long prompts you inherited | You must review its output |
None of these cost anything beyond the usage you already pay for. If you are on a paid plan, check the £ price on Anthropic's pricing page at checkout, since that changes.
Take each prohibition and ask "what should it do instead?" Write that.
| Prohibition | Swap |
|---|---|
| Don't be wordy. | Answer in no more than five sentences, leading with the answer. |
| Don't use bullet points. | Write in plain paragraphs, with the reasoning running as a connected argument. |
| Don't make things up. | Use only the facts in the document I paste. If it does not say, write "not stated". |
| Don't sound like a robot. | Write as a colleague would in an email: first person, short sentences, one contraction per paragraph at most. |
| Don't give generic advice. | Give three options specific to [MY_SITUATION], each with a first step I can do today. |
Notice the swaps are testable. You can check whether an answer is five sentences long. You cannot easily check whether it is "not wordy".
After a swap, add one clause saying why. "Keep replies under 120 words, because I read these on my phone between meetings." That tells Claude what to trade off when a case comes up that your rule did not foresee, for instance whether to drop a caveat or a second example.
If you cannot write an honest reason, the rule may be a habit rather than a need. Delete it and see whether the output suffers.
Some prohibitions are right. "Do not send anything; draft only" and "Do not include personal data from the attached file in the summary" are limits, not style notes. Keep them, but make them narrow and give them a reason: "Draft only, because I review everything before it goes to a client." Put them in one short block near the end, so they read as limits and not as the main character of the prompt.
Paste your existing prompt in and have Claude do the first pass of swaps. Fill in the bracketed parts.
You are an editor who specialises in writing clear instructions for AI assistants.
Below is a prompt I currently use. It contains many negative instructions (things to avoid).
<my_prompt>
[PASTE_YOUR_PROMPT]
</my_prompt>
What the prompt is for: [PURPOSE]
Who reads the output and where: [AUDIENCE_AND_PLACE]
What a good result looks like to me: [SUCCESS_DESCRIPTION]
Steps:
1. List every negative instruction (anything phrased as don't, never, avoid, do not, without).
2. For each one, work out what I most likely want instead. If it is unclear, put it in a "Questions for me" list rather than guessing.
3. Rewrite each as a positive, checkable instruction.
4. Where a reason is obvious from my purpose, add it as a short "because" clause. Do not invent reasons.
5. Mark any prohibition that guards a genuine limit (accuracy, privacy, sending, deleting) as KEEP, and rewrite it to be specific, with a reason.
6. Return the full revised prompt, then a two-column table of old line and new line.
Constraints: preserve my intent, keep my wording where it already works, no extra rules I did not imply, British English.
Before answering, check that no positive instruction contradicts another, and that each KEEP line is genuinely a limit.
Fill in: your prompt, its purpose, the audience, and what success looks like.
Rewrites are a hypothesis. Run the old prompt and the new one on the same three realistic inputs, and compare against the list of things you actually care about. A swap that is more precise but changes the content you wanted has not improved anything. Use this prompt to compare:
You are a careful reviewer. I will give you one task, two versions of a prompt, and the output each produced.
Task: [TASK]
What matters to me, in order: [CRITERIA]
<prompt_a_and_output>[OLD_PROMPT_AND_OUTPUT]</prompt_a_and_output>
<prompt_b_and_output>[NEW_PROMPT_AND_OUTPUT]</prompt_b_and_output>
Score each output against my criteria with one sentence of evidence from the text. Say which is better and why. If the difference is small or one example cannot settle it, say so rather than picking a winner.