The five mistakes
- No length cap. 'Write a post about invoicing' produces a post of random size. 'Four sentences' produces four sentences.
- No audience named. The model writes for everyone, which lands with no one. 'For freelance designers' changes every word that follows.
- No example. One sample of the tone you want is worth three paragraphs describing it.
- Too many jobs in one prompt. Write it, improve it, translate it, summarize it - four prompts, in order, each seeing the last one's output.
- Vague adjectives instead of constraints. 'Make it engaging and punchy' does nothing; 'ban the words unlock, elevate and game-changing' does something measurable.
Fixing a prompt, step by step
- State the job in one sentence at the top, with the audience: 'Explain to a small business owner why...'
- Cap the output: a sentence count, a word count, a line count. Any cap beats no cap.
- Add the format by example: paste two lines you like and say 'like this'.
- List the bans: words to avoid, formats to skip, no emoji, no hashtags - whatever your taste actually is.
- Iterate in follow-ups, not in one giant prompt: 'shorter', 'keep the first line, redo the rest', 'now the same for plumbers'.
The test that tells you a prompt is done
Run the finished prompt three times on three different days of the week, or with three different starting inputs. If the outputs vary wildly in usefulness, the prompt is still under-specified; if they vary in flavor but not in quality, it is done. This is also the honest argument for keeping prompts in a file rather than in chat history: a prompt you cannot re-run identically is a memory, not a tool. The vault metaphor is literal - stored, versioned, tested, reused.
- A prompt is a tool when it produces usable output on an ordinary day without you in the room.
- Version prompts like code: when one improves, overwrite deliberately, not accidentally.
- The prompt library should outlive any single chat account, model or price change.
Small rules with outsized effect
- Put the constraints at the end of the prompt, not the start - instructions near the output shape the output.
- Ask for options, then choose: 'give me three subject lines' beats hoping the first is right.
- Tell it what to do with uncertainty: 'if you do not know, say so' prevents confident nonsense.
- Save the winners. The ten prompts you actually reuse are the entire library; everything else is clutter.
The free tool for this
There is no calculator that answers this one honestly, so here is the free set: eight calculators for rates, quotes, invoices, splits, prices and fees. They run in the page and send nothing anywhere.
If you would rather not build it yourself
The Prompt Vault is 79 prompts built this way - length capped, audience named, buzzwords banned - and the free sampler is real pages from it, no email wall. Prompt Vault is $12. See it on the shop, and there is a plain page about it here: what it holds.
Questions people actually ask
Do long prompts work better than short ones?
Neither length wins - specificity wins. A short prompt with a hard cap and a named audience beats a long one that describes the task in five airy ways.
Should I give the model a role, like 'act as a senior copywriter'?
Roles can help tone, but they are no substitute for constraints. 'Act as' plus a length cap plus an example beats 'act as' alone, every time it is tested.
Why do my prompts work one week and fail the next?
Because they were under-specified and the first success was luck. The three-runs test catches this: a prompt that only worked once was never really working.
Read next
How to write a better AI prompt for real work, Where to keep the AI prompts that work, so you stop rewriting them.