Why this matters

Custom instructions are standing orders the assistant reads before every reply. Most bad output is a briefing problem, and ten lines fix nine complaints out of ten.

The 15-minute experiment

1. 1Write down five things it keeps getting wrong, in your own words: 'too long', 'uses the word delve', 'too American'
2. 2Turn each gripe into an instruction under 12 words: 'Keep replies under 150 words'
3. 3In ChatGPT: Settings, then Personalisation, then Custom instructions. In Claude: Settings, then Profile
4. 4Paste them in, save, then re-run a prompt you used yesterday

Rewrite these five complaints as short, testable instructions for an assistant. Keep them under 12 words each.

• Five gripes written down
• Each rewritten as one short rule
• Rules saved in settings and still there after a refresh
• Old prompt re-run and at least two gripes gone