Best practice L2 · Context engineering informational

Tune Custom Instructions From Answers You Didn't Like

What it is

A maintenance habit: treat every correction you find yourself repeating ('shorter', 'no bullet points', 'British spelling') as a signal to add a permanent custom instruction, so you fix it once instead of every chat.

let annoyances drive editsNotice a recurring annoyanceWrite it as a ruleAdd to custom instructions
Every repeated correction is a custom-instruction waiting to be written.

Why it works

Custom instructions work best when they're grown from real friction, not written in one guess up front. The things that annoy you repeatedly are exactly the defaults worth changing. Converting each recurring correction into a rule steadily shrinks the gap between the first answer and the one you actually want.

When to use it

Whenever you catch yourself giving the same follow-up correction for the third time. Do a periodic sweep of your instructions every month or so.

When not to use it

One-off preferences that only applied to a single task — encoding those as permanent rules makes ChatGPT rigid where you wanted it flexible.

Prompt

I keep having to tell you: <the correction you repeat>. Rewrite my custom instructions to bake this in permanently, without over-constraining other tasks. Give me the exact text to paste.

Example

After correcting 'too formal' for the fifth time, you add 'write like you're explaining to a smart friend' — and stop fighting the tone in every thread.

Common mistakes

  • Re-typing the same correction forever instead of promoting it to a rule.
  • Piling in so many rules that answers become stilted and over-hedged.
  • Never revisiting instructions, so stale rules from an old project still fire.

Related