Tips

Prompt Tips Tested on Claude, GPT and Gemini

I got tired of hunting for prompting tips that actually work, so I built a place to collect and test them.

18 tips tested 10 models 4,938 answers scored

17 tips, from 18 tested claims: the examples family reports two claims on one page.

How to read this

Each tip rests on a claim tested as a paired comparison on a fixed task set. Across 54 tested pairs the verdicts are 23 In free drift, 16 Stable, 14 Unobservable, 1 Past the horizon.

In free drift: could not be told from no effect; Stable: held; Unobservable: the test could not measure it; Past the horizon: measured worse.

Filter tips

Showing 17 of 17 tips.

What the marks mean
- Holds
- No measured effect
- Could not measure
- Helped a little, under our bar
- Scored worse
Do this
7 of 7Held on at least one model, with the models named where it did not.
Holds
Stop made-up answerslead
Tell the model it is allowed to say it does not know. Make that an acceptable answer.
How to use itAdd a line saying that if the text does not answer the question, replying I do not know is fine. Put it next to your question.
Pass rate up 80 to 100 points on all three models.
Models helped
Via API · Claude Haiku 4.5 · Holds · GPT-5 mini · Holds · Gemini 3.1 Flash Lite · Holds
In Claude Code · Claude Fable 5 · Holds · Claude Opus 5 · Holds · Claude Haiku 4.5 · Holds · Claude Sonnet 5 · Holds · Claude Fable 5.1 · Holds
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Holds · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Holds · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Holds
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Holds
Better step-by-step answersco-lead
Ask the model to work through the problem step by step before it answers.
How to use itAdd a line asking for the steps before the answer on any task with chained arithmetic or logic. Check that it helps on the model you use.
Pass rate up 20 points on Claude and 34 points on GPT. Inside the margin on Gemini.
Models helped
Via API · Claude Haiku 4.5 · Holds · GPT-5 mini · Holds · Gemini 3.1 Flash Lite · No measured effect
In Claude Code · Claude Fable 5 · No measured effect · Claude Opus 5 · No measured effect · Claude Haiku 4.5 · Holds · Claude Sonnet 5 · No measured effect · Claude Fable 5.1 · No measured effect
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · No measured effect · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · No measured effect · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · No measured effect
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Holds
Getting clean JSON back
Show the model the exact shape you want back. Write out the field names instead of just asking for JSON.
How to use itPut a small example of the structure in your prompt. List every field you need and what kind of value it holds.
Pass rate rose from 0% to 100% on all three models.
Models helped
Via API · Claude Haiku 4.5 · Holds · GPT-5 mini · Holds · Gemini 3.1 Flash Lite · Holds
In Claude Code · Claude Fable 5 · Holds · Claude Opus 5 · Holds · Claude Haiku 4.5 · Holds · Claude Sonnet 5 · Holds · Claude Fable 5.1 · Holds
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Holds · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Holds · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Holds
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Holds
Showing examples
Give one good example of the answer you want. A second or third adds very little.
How to use itPaste one example in the format you need. Add more only for formats that are tricky to describe in words. This helped a little on 9 of 10 models tested, under our bar for a clear win.
Pass rate up 18 to 19 points on all three models.
Models helped
Via API · Claude Haiku 4.5 · Helped a little, under our bar · GPT-5 mini · Helped a little, under our bar · Gemini 3.1 Flash Lite · Helped a little, under our bar · Claude Haiku 4.5 · Helped a little, under our bar · GPT-5 mini · Helped a little, under our bar · Gemini 3.1 Flash Lite · Helped a little, under our bar
In Claude Code · Claude Fable 5 · Could not measure · Claude Opus 5 · Could not measure · Claude Fable 5 · Helped a little, under our bar · Claude Opus 5 · Helped a little, under our bar · Claude Fable 5 · Helped a little, under our bar · Claude Opus 5 · Helped a little, under our bar · Claude Haiku 4.5 · Could not measure · Claude Sonnet 5 · Could not measure · Claude Haiku 4.5 · Helped a little, under our bar · Claude Sonnet 5 · Helped a little, under our bar · Claude Haiku 4.5 · Helped a little, under our bar · Claude Sonnet 5 · Helped a little, under our bar · Claude Fable 5.1 · Could not measure · Claude Fable 5.1 · Helped a little, under our bar · Claude Fable 5.1 · Helped a little, under our bar
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Could not measure · GPT-5.4 mini · Codex run on GPT-5.4 mini · Holds · GPT-5.4 mini · Codex run on GPT-5.4 mini · Holds · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Could not measure · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Helped a little, under our bar · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Helped a little, under our bar · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Could not measure · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Helped a little, under our bar · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Helped a little, under our bar
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Holds
Asking for a rewrite
Ask for a draft, then ask the model to criticise it and rewrite.
How to use itIn one prompt, ask for a draft, then a short critique, then a final version. Compare it against what a single pass gives you.
Rubric mean up 0.30 points on Claude. Inside the margin on GPT and Gemini.
Models helped
Via API · Claude Haiku 4.5 · Holds · GPT-5 mini · No measured effect · Gemini 3.1 Flash Lite · No measured effect
One square is one model, on one test method. Models we did not test on this tip have no square.
Show the per-model numbers
Holds
Splitting big tasks
Break a big task into a few prompts, one step at a time.
How to use itAsk for the first step, then feed its answer into the next prompt. Best on tasks with several parts. Check that it helps on the model you use.
Pass rate up 10 to 25 points on all three models.
Models helped
Via API · Claude Haiku 4.5 · Holds · GPT-5 mini · Holds · Gemini 3.1 Flash Lite · No measured effect
In Claude Code · Claude Haiku 4.5 · multi-turn · Holds
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Holds
Getting the length right
Ask for an exact word count. Vague words like short miss by a lot.
How to use itSay exactly how many words you want rather than keep it short. Expect the answer to land within a few percent of the number you asked for.
Pass rate up 68 to 95 points on all three models.
Models helped
Via API · Claude Haiku 4.5 · Holds · GPT-5 mini · Holds · Gemini 3.1 Flash Lite · Holds · GPT-5.4 mini · API twin for the Codex run · Holds
In Claude Code · Claude Fable 5 · Holds · Claude Opus 5 · Holds · Claude Haiku 4.5 · Holds · Claude Sonnet 5 · Holds · Claude Fable 5.1 · Holds
In Codex · GPT-5.4 mini · Codex run, replaying two earlier tests · Holds · GPT-5.4 mini · Codex run on GPT-5.4 mini · Holds · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Holds · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Holds
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Skip this
6 of 6Widely repeated, and showed no measured effect on any model tested.
Debunked
Giving the model a job title
You have probably heard you should tell the model to act as an expert.
Giving the model a job title did not change answer quality in our tests. Write a clearer question instead.
What to do insteadSpend those words on the task. Say what you want, who will read it, and what a good answer has to cover.
Rubric mean moved at most 0.13 points on all three models, inside the margin every time.
Models helped
Via API · Claude Haiku 4.5 · No measured effect · GPT-5 mini · No measured effect · Gemini 3.1 Flash Lite · No measured effect
One square is one model, on one test method. Models we did not test on this tip have no square.
Show the per-model numbers
Debunked
Adding pressure
You have probably heard that telling the model your job depends on it gets you a better answer.
Adding pressure or high stakes to your prompt did not change answer quality in our tests.
What to do insteadDescribe the task and the standard you want. Concrete requirements do the work that urgency does not.
Rubric mean moved at most 0.11 points on all three models, inside the margin every time.
Models helped
Via API · Claude Haiku 4.5 · No measured effect · GPT-5 mini · No measured effect · Gemini 3.1 Flash Lite · No measured effect
One square is one model, on one test method. Models we did not test on this tip have no square.
Show the per-model numbers
Debunked
Being polite
You have probably heard that being polite, or being blunt, changes what you get back.
Please and thank you did not change answer quality in our tests. Write however you prefer.
What to do insteadBe as plain or as warm as you like. Spend the effort on stating the task clearly.
Rubric mean moved at most 0.13 points on all three models, inside the margin every time.
Models helped
Via API · Claude Haiku 4.5 · No measured effect · GPT-5 mini · No measured effect · Gemini 3.1 Flash Lite · No measured effect
One square is one model, on one test method. Models we did not test on this tip have no square.
Show the per-model numbers
Debunked
Rules for the whole chat
You have probably heard that rules in the system prompt stick better than rules in the message.
Where you put your standing rules did not change how well the model kept them.
What to do insteadPut rules wherever is convenient, and repeat them if the chat runs long.
Measured on two models; the rest could not be measured on this task set.
Models helped
Via API · Claude Haiku 4.5 · Could not measure · GPT-5 mini · No measured effect · Gemini 3.1 Flash Lite · No measured effect
In Claude Code · Claude Haiku 4.5 · multi-turn · Could not measure
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Debunked
Prompt, then revise
You have probably heard that asking the model to critique and revise its own draft improves it.
Asking the model to revise its own draft did not make it better.
What to do insteadWrite one careful prompt. Revise only when you can name what is wrong.
Measured on two models; the rest could not be measured on this task set.
Models helped
Via API · Claude Haiku 4.5 · Could not measure · GPT-5 mini · No measured effect · Gemini 3.1 Flash Lite · No measured effect
In Claude Code · Claude Haiku 4.5 · multi-turn · No measured effect
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Debunked
Questions about recent events
You have probably heard that telling the model its knowledge may be out of date makes it more careful.
Telling the model its knowledge may be out of date makes it hedge more.
What to do insteadOnly add it for questions about recent events, and check the marks: on Claude it made answers worse.
Pass rate up 41 points on GPT and 30 points on Gemini, and scored worse on Claude by 23 points.
Models helped
Via API · Claude Haiku 4.5 · Scored worse · GPT-5 mini · Holds · Gemini 3.1 Flash Lite · Holds · GPT-5.4 mini · API twin for the Codex run · Holds
In Claude Code · Claude Fable 5 · No measured effect · Claude Opus 5 · No measured effect · Claude Haiku 4.5 · No measured effect · Claude Sonnet 5 · No measured effect · Claude Fable 5.1 · No measured effect
In Codex · GPT-5.4 mini · Codex run, replaying two earlier tests · Holds · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · No measured effect
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Inconclusive
4 of 4The instrument could not resolve these. No application claim is made.
Inconclusive
Instructions before or after
We could not measure this one. Our task was easy enough that both orderings scored full marks, so nothing separated them.
Every model scored full marks with and without it, so the task set separated nothing.
Models helped
Via API · Claude Haiku 4.5 · Could not measure · GPT-5 mini · Could not measure · Gemini 3.1 Flash Lite · Could not measure
In Claude Code · Claude Fable 5 · Could not measure · Claude Opus 5 · Could not measure · Claude Haiku 4.5 · Could not measure · Claude Sonnet 5 · Could not measure · Claude Fable 5.1 · Could not measure
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Could not measure · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Could not measure · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Could not measure
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Inconclusive
Using tags and formatting
We could not measure this one. Both ways of marking up the prompt scored full marks, so the test could not tell them apart.
Every model scored full marks with and without it, so the task set separated nothing.
Models helped
Via API · Claude Haiku 4.5 · Could not measure · GPT-5 mini · Could not measure · Gemini 3.1 Flash Lite · Could not measure
In Claude Code · Claude Fable 5 · Could not measure · Claude Opus 5 · Could not measure · Claude Haiku 4.5 · Could not measure · Claude Sonnet 5 · Could not measure · Claude Fable 5.1 · Could not measure
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Could not measure · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · No measured effect · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Could not measure
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Inconclusive
Instructions in long prompts
We could not measure this one. The models followed the instruction wherever we put it, so the test could not separate the two.
Every model scored full marks with and without it, so the task set separated nothing.
Models helped
Via API · Claude Haiku 4.5 · Could not measure · GPT-5 mini · Could not measure · Gemini 3.1 Flash Lite · Could not measure
In Claude Code · Claude Fable 5 · Could not measure · Claude Opus 5 · Could not measure · Claude Haiku 4.5 · Could not measure · Claude Sonnet 5 · Could not measure · Claude Fable 5.1 · Could not measure
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Could not measure · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Could not measure · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Could not measure
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers
Inconclusive
Offering the model money
Offering the model a tip made answers longer. We could not measure whether they got better, which is what the claim is really about.
Answer length rose by 44 to 180 words, but this metric has no threshold, so no verdict follows.
Models helped
Via API · Claude Haiku 4.5 · Could not measure · GPT-5 mini · Could not measure · Gemini 3.1 Flash Lite · Could not measure
In Claude Code · Claude Fable 5 · Could not measure · Claude Opus 5 · Could not measure · Claude Haiku 4.5 · Could not measure · Claude Sonnet 5 · Could not measure · Claude Fable 5.1 · Could not measure
In Codex · GPT-5.4 mini · Codex run on GPT-5.4 mini · Could not measure · GPT-5.6 Luna · Codex run on GPT-5.6 Luna · Could not measure · GPT-5.6 Terra · Codex run on GPT-5.6 Terra · Could not measure
One square is one model, on one test method. A named run is a second reading of the same model. Models we did not test on this tip have no square.
Show the per-model numbers

Showing 17 of 17 tips.

Every verdict, by model

One row per tip, one column per model, on the instrument that measured it. The two instruments are two tables and are never read as one set: the same model on the API and in Claude Code is two measurements, not two opinions.

Tips · tested via API
Tip	Claude Haiku 4.5 ·	GPT-5 mini ·	Gemini 3.1 Flash Lite ·	GPT-5.4 mini · API twin for the Codex run ·
Getting clean JSON back	Holds · +1.000	Holds · +1.000	Holds · +1.000	Not run
Instructions before or after	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Using tags and formatting	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Showing examples	No measured effect · +0.157	No measured effect · +0.186	No measured effect · +0.157	Not run
Better step-by-step answers	Holds · +0.200	Holds · +0.340	No measured effect · +0.100	Not run
Stop made-up answers	Holds · +1.000	Holds · +0.800	Holds · +1.000	Not run
Instructions in long prompts	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Offering the model money	Could not measure · +44.2	Could not measure · +179.6	Could not measure · +56.5	Not run
Giving the model a job title	No measured effect · -0.013	No measured effect · +0.133	No measured effect · -0.033	Not run
Adding pressure	No measured effect · +0.033	No measured effect · +0.111	No measured effect · -0.011	Not run
Being polite	No measured effect · -0.100	No measured effect · -0.056	No measured effect · -0.128	Not run
Asking for a rewrite	Holds · +0.300	No measured effect · +0.022	No measured effect · -0.067	Not run
Rules for the whole chat	Could not measure · +0.000	No measured effect · +0.000	No measured effect · +0.100	Not run
Splitting big tasks	Holds · +0.233	Holds · +0.247	No measured effect · +0.100	Not run
Prompt, then revise	Could not measure · +0.000	No measured effect · +0.029	No measured effect · +0.000	Not run
Questions about recent events	Scored worse · -0.233	Holds · +0.411	Holds · +0.300	Holds · +0.592
Getting the length right	Holds · +0.701	Holds · +0.952	Holds · +0.684	Holds · +0.812

Tips · tested in Claude Code
Tip	Claude Fable 5 ·	Claude Opus 5 ·	Claude Haiku 4.5 ·	Claude Sonnet 5 ·	Claude Fable 5.1 ·	Claude Haiku 4.5 · multi-turn ·
Getting clean JSON back	Holds · +1.000	Holds · +1.000	Holds · +1.000	Holds · +1.000	Holds · +1.000	Not run
Instructions before or after	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Using tags and formatting	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Showing examples	No measured effect · +0.180	No measured effect · +0.183	No measured effect · +0.174	No measured effect · +0.183	No measured effect · +0.186	Not run
Better step-by-step answers	No measured effect · +0.100	No measured effect · +0.060	Holds · +0.300	No measured effect · +0.100	No measured effect · +0.100	Not run
Stop made-up answers	Holds · +0.760	Holds · +0.700	Holds · +0.780	Holds · +0.620	Holds · +0.240	Not run
Instructions in long prompts	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Offering the model money	Could not measure · +62.0	Could not measure · +283.5	Could not measure · +16.6	Could not measure · +80.0	Could not measure · +105.1	Not run
Giving the model a job title	Not run	Not run	Not run	Not run	Not run	Not run
Adding pressure	Not run	Not run	Not run	Not run	Not run	Not run
Being polite	Not run	Not run	Not run	Not run	Not run	Not run
Asking for a rewrite	Not run	Not run	Not run	Not run	Not run	Not run
Rules for the whole chat	Not run	Not run	Not run	Not run	Not run	Could not measure · +0.000
Splitting big tasks	Not run	Not run	Not run	Not run	Not run	Holds · +0.313
Prompt, then revise	Not run	Not run	Not run	Not run	Not run	No measured effect · +0.029
Questions about recent events	No measured effect · +0.078	No measured effect · +0.064	No measured effect · +0.025	No measured effect · -0.008	No measured effect · -0.031	Not run
Getting the length right	Holds · +0.235	Holds · +0.823	Holds · +0.650	Holds · +0.774	Holds · +0.494	Not run

Tips · tested in Codex
Tip	GPT-5.4 mini · Codex run on GPT-5.4 mini ·	GPT-5.6 Luna · Codex run on GPT-5.6 Luna ·	GPT-5.6 Terra · Codex run on GPT-5.6 Terra ·	GPT-5.4 mini · Codex run, replaying two earlier tests ·
Getting clean JSON back	Holds · +1.000	Holds · +1.000	Holds · +1.000	Not run
Instructions before or after	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Using tags and formatting	Could not measure · +0.000	No measured effect · -0.100	Could not measure · +0.000	Not run
Showing examples	Holds · +0.200	No measured effect · +0.129	No measured effect · +0.186	Not run
Better step-by-step answers	No measured effect · +0.100	No measured effect · +0.100	No measured effect · +0.100	Not run
Stop made-up answers	Holds · +0.600	Holds · +0.700	Holds · +0.800	Not run
Instructions in long prompts	Could not measure · +0.000	Could not measure · +0.000	Could not measure · +0.000	Not run
Offering the model money	Could not measure · +102.9	Could not measure · +83.5	Could not measure · +82.1	Not run
Giving the model a job title	Not run	Not run	Not run	Not run
Adding pressure	Not run	Not run	Not run	Not run
Being polite	Not run	Not run	Not run	Not run
Asking for a rewrite	Not run	Not run	Not run	Not run
Rules for the whole chat	Not run	Not run	Not run	Not run
Splitting big tasks	Not run	Not run	Not run	Not run
Prompt, then revise	Not run	Not run	Not run	Not run
Questions about recent events	Not run	Not run	No measured effect · +0.100	Holds · +0.542
Getting the length right	Holds · +0.748	Holds · +0.369	Holds · +0.290	Holds · +0.781

Each cell is one measured pair: the verdict and the delta that cell holds. Nothing on this grid is averaged, across models, instruments or runs. Sorting orders rows by a column's delta; cells with no delta sort last in both directions. Fable 5 and Fable 5.1 are separate columns: they are two versions and this site does not merge them.

What the marks mean

Do this

Skip this

Inconclusive

Every verdict, by model