How to write a prompt that works the second time you run it
A prompting guide built on what changes the output and what only looks like it does, with a six block skeleton and the test that catches a confident wrong answer.
- Works in
- Claude
- You need
- A job you run more than once · One example of a good result
- Written for
- prompting guide ai
Scored by our own engine
This page, run through the audit we sell. Measured 11 August 2026.
Anyone can get a good answer out of an assistant once. The useful skill is writing something that produces the same quality on Thursday, from a colleague, on an input you have not seen. That takes six blocks, one example and two tests.
The four things that carry most of the effect
Name the deliverable, not the topic. Almost every unreliable prompt fails here. Competitor analysis gets you an essay. A table of six competitors with columns for pricing model, entry price, free tier and the one thing they say they do better gets you the same artefact every time.
Give one example of done. A single example encodes length, order, register and depth simultaneously, and it does it better than describing all four. This is the cheapest large improvement available and the block people skip most.
Specify the format exactly. Columns, section names, JSON keys, a maximum length, and the instruction to output nothing before or after it. Left unspecified, the model chooses, and it chooses differently under a slightly different input.
Say what to do when something is missing. Without this, a gap in the input is filled with something plausible. With it, the gap comes back as a marked blank. This single line is the difference between an assistant that helps and one that quietly manufactures your evidence.
What does less than people think
Role framing. You are a world class SEO strategist changes the vocabulary slightly and the reasoning not at all. Where a role genuinely changes audience or register, writing for clinicians rather than patients, keep it. Otherwise it is context spent on nothing.
Think step by step. It earned its reputation honestly: in the 2022 paper that introduced it, adding the phrase lifted text-davinci-002 from 17.7% to 78.7% on MultiArith and from 10.4% to 40.7% on GSM8K. Current models with built in reasoning already do this, so on most tasks today the phrase adds words rather than accuracy. It still earns its place on multi step arithmetic and logic, particularly on smaller models.
Politeness, tips and threats. Please, a promised bonus, a warning about consequences. Reported effects are small, inconsistent across models and unstable across versions. A prompt that works because of one of these is a prompt that will stop working after an update, and you will not know why.
Piling on constraints. Every rule competes for attention with the task. The symptom of too many is that the specific ones stop being followed while the general tone survives, which reads like the model getting worse and is actually the prompt getting fatter.
The two tests
Run it twice on the same input, in separate chats. Whatever differs between the two runs is what you did not specify. That is the whole diagnosis, and it takes two minutes.
Run it once on an input that does not contain the answer. Delete the key paragraph from the source material and see what happens. The correct response is to say the material does not cover it. The common response is a fluent answer with the missing piece invented.
We ran that second test deliberately on a research prompt with web access switched off. One assistant said plainly that it could not browse or verify anything. Another produced the full report with a tidy citation list, in which one URL was a hard 404. The honest output looked worse. That is why this test belongs in the routine rather than in an appendix.
Where prompts belong once they work
A prompt that works and lives in somebody’s chat history is not an asset. Keep it with its variables and its example, so the next person to run it knows what goes in the placeholders. Our prompt library is that store, published rather than internal, and every page in it carries the inputs and the failure mode alongside the prompt.
If you are starting from a blank box, the prompt generator builds the six blocks from a description of the job. If you already have one that half works, the prompt optimizer scores it and names the single cause. And if the prompt is producing pages you publish, check the page rather than the prompt: the AI content readiness checker tells you whether anything on it can actually be lifted and quoted.
Source
- Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo and Yusuke Iwasawa, Large Language Models are Zero-Shot Reasoners, arXiv:2205.11916, 24 May 2022, for the MultiArith and GSM8K figures quoted above. https://arxiv.org/abs/2205.11916
## 1. Deliverable
Produce [WHAT SHOULD EXIST WHEN THIS IS FINISHED]. It is for
[WHO OR WHAT CONSUMES IT].
## 2. Inputs
[PASTE THE MATERIAL, OR NAME THE PLACEHOLDERS]
Use only what is above. If something needed is missing, say which and
continue with the rest.
## 3. Method
[THE STEPS, NUMBERED, ONLY WHERE ORDER CHANGES THE RESULT]
## 4. Output format
[THE EXACT SHAPE: TABLE WITH THESE COLUMNS, LIST OF N, DOCUMENT WITH THESE
SECTIONS, JSON WITH THESE KEYS]
Maximum length: [LIMIT]. Nothing before it and nothing after it.
## 5. Constraints
- Do not invent figures, statistics, sources, quotes or examples. If a claim
needs a number you do not have, write the claim without it or mark it.
- Do not [THE FAILURE YOU HAVE ALREADY SEEN THIS JOB PRODUCE].
- State assumptions, labelled, and continue. Do not ask questions first.
## 6. Done looks like
[ONE EXAMPLE OF A GOOD OUTPUT, OR THE OPENING TWO LINES OF ONE]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[WHAT SHOULD EXIST WHEN THIS IS FINISHED]- The artefact, named. A table of twelve keywords grouped by intent is a deliverable. Keyword research is a topic, and a topic gets you whatever the model considers a reasonable response to a topic, which changes every run.
[WHO OR WHAT CONSUMES IT]- A person reading it in an email, a developer reading it in a ticket, or a script parsing it. This single line decides tone, length and format together, and leaving it out is why output arrives in the wrong register and gets rewritten by hand every time.
[THE STEPS, NUMBERED, ONLY WHERE ORDER CHANGES THE RESULT]- Steps for genuinely sequential work: classify then group then rank. Do not number things that are not sequential. A model given six ceremonial steps spends effort satisfying the ceremony rather than the task.
[THE FAILURE YOU HAVE ALREADY SEEN THIS JOB PRODUCE]- One observed failure, not a list of hypothetical ones. Do not pad the summary to fill the word count. Do not treat the caveat as optional. Each constraint you add costs attention, so they should be earned by something that actually went wrong.
[ONE EXAMPLE OF A GOOD OUTPUT, OR THE OPENING TWO LINES OF ONE]- The highest value block in the skeleton and the one people skip. An example fixes length, order, register and depth in one move, and it does it more reliably than any description of those four things.
How to run it
- 01Write the deliverable line first
Before anything else, finish the sentence: when this prompt has run, there will be a ___. If you cannot finish it, the prompt is not ready and no amount of instruction quality will rescue it.
- 02Paste an example of done
One real example beats three paragraphs describing what you want. Use an imperfect one and say which part of it is good. This is the difference between a prompt that produces your format and one that produces the format the model considers standard.
- 03Run it twice on the same input, in two fresh chats
A prompt that works once has told you nothing. Two independent runs on the same input show you the variance, and the parts that differ are exactly the parts you failed to specify. Fix those and nothing else.
- 04Run it once on an input that does not contain the answer
This is the test that matters. Give it material with a gap in it and see whether it says so or fills it in. A prompt that fills it in will do the same thing on real data, fluently, and you will not notice.
- 05Delete a constraint every time you add one
Prompts grow after every bad output until they are longer than the work and the model is averaging twenty rules instead of following five. When you add a rule, read the list and remove one that has not earned its place.
Questions people ask
What actually makes a prompt work?
Four things, in order of effect: naming the deliverable rather than the topic, giving one example of a good output, specifying the output format exactly, and stating what to do when the input is missing something. Everything else is refinement. A prompt with those four is reliable enough to hand to somebody else.
Does saying "think step by step" still help?
Much less than it did, and on some models it is now redundant. The phrase comes from a 2022 paper where it raised accuracy on MultiArith from 17.7% to 78.7% and on GSM8K from 10.4% to 40.7% for text-davinci-002. Current reasoning models already decompose problems by default, so the phrase mostly adds length. Keep it for genuinely multi step arithmetic or logic on smaller or older models, and drop it from writing tasks.
Do politeness, tips or threats change the output?
Not in any way you should build on. Please and thank you cost nothing and do nothing. Offering an imaginary tip or threatening a consequence produces inconsistent effects across models and versions, and a prompt whose performance depends on a superstition breaks silently when the model is updated. Spend those tokens on an example instead.
How long should a prompt be?
As long as its inputs need and as short as its rules allow. There is no target word count, but there is a reliable symptom of being too long: the specific rules stop being followed while the general tone survives. When that happens, cut constraints that were added for failures that happened once.
Should I use one big prompt or several small ones?
Split when the output of one part is something you would want to check before continuing. Research then draft is two prompts because the research is worth reading before it becomes prose. Classify then group is one prompt because you would never inspect the classification on its own.
How do I stop a model inventing statistics?
Forbid it explicitly and give it somewhere to put the gap, for example a marker like NEEDS FIGURE. A model produces an invented number because a sentence with a hole in it looks like a failed answer, so the fix is to make the hole an acceptable output. Then check anything that survives against a source you can open.
Unsubscribe in one click. We never pass your address on.