The prompt that fixes your other prompts

Here is a prompt I use:

{paste your prompt}

Critique it before you improve it:
1. List every place a model could reasonably read an instruction two
   different ways.
2. List instructions that conflict with each other, and instructions that
   are wishes rather than rules (things a model can quietly ignore with no
   visible failure).
3. List anything I am clearly expecting but never actually asked for.

Then rewrite it. Keep it as short as the job allows. Every instruction in
the rewrite must be checkable: I should be able to look at any output and
say whether the instruction was followed or not.

End with two test inputs I can run through both versions to compare them.

Prompts rot the same way shell scripts do. You write one, it mostly works, you bolt on another rule every time it misbehaves, and a month later it is twelve instructions long and the model ignores four of them. I point this at any prompt I have patched more than twice, including the ones on this site.

Why it works

• Critique before rewrite, in that order. Ask for an improved prompt directly and you get cosmetic edits with confident commentary. Making the model name specific ambiguities first forces it to actually read the thing, and the rewrite quality follows.
• “Wishes rather than rules” is the category that earns its keep. “Be concise” is a wish. “Under 150 words” is a rule. Most tired prompts are mostly wishes, and seeing them listed is usually the whole diagnosis.
• The checkability requirement keeps the rewrite honest too. It stops the model from replacing your vague instructions with new, better-sounding vague instructions, which is exactly what it wants to do.
• The test inputs make it falsifiable. Two runs through old and new, and you know whether this was an improvement or a lateral move. Skip this step and you are just collecting prompts that feel better.

What good output looks like

The ambiguity list should sting a little. When I ran my own news-triage prompt through this, it flagged that I asked for things being “downplayed” without ever defining what counts, so the model was free to decide, and it decided differently every run. One definition later, the output stabilized.

Where it works (and doesn’t)

Claude and GPT-class models only. This is a reasoning-about-instructions task, and local models at the 14B level produce generic advice (“add more detail”) instead of specific findings. One habit worth keeping: reject any rewrite that is longer than your original. Length is where prompt rot starts, and a fixer that adds words is reinfecting the patient.

The uncomfortable lesson after enough runs: the model was never ignoring your instructions. You just never actually gave them.