Ask better

Structure beats clever wording. Give context, constraints, and what good looks like.

The shape of a strong ask

Clever wording is optional. Structure is not.

Role or lens - who should the model act like, briefly.
Context - audience, tools, constraints, what you already tried.
Task - one primary verb.
Output form - bullets, table, code, checklist, length.
Quality bar - examples of good and bad, or acceptance criteria.

If you only remember three words: context, constraints, criteria.

Template you can reuse

You are helping a [role] who needs [outcome].
Context: [audience, tools, constraints, what I already tried].
Task: [one verb] [specific deliverable].
Output: [format, length, sections].
Quality bar: [what good looks like]. Avoid: [what bad looks like].
If unsure, list assumptions instead of inventing facts.

Fill every bracket. Delete lines you do not need. Do not add five tasks in one ask.

Before / after

Before (weak):

Make this email better.

After (strong):

You are helping a busy operations lead.
Context: 12-person MSP; internal status email to a client contact; no PHI; tone calm and specific.
Task: Revise the draft below into 6 short bullets plus one clear ask.
Output: Markdown. Flag any claim that needs a human fact-check.
Quality bar: Concrete dates and owners. Avoid: hype, vague "synergy," invented policy.
Draft:
[paste]

The second version steers the model and tells you what to verify.

Example skeleton (ops)

You are helping a busy operations lead.
Context: We run a 12-person MSP; no PHI in this draft.
Task: Draft a 6-bullet checklist for verifying an AI-written client email.
Output: Markdown checklist. Flag anything that needs a human.

Anti-patterns

Dumping a vague “make this better” with no goal.
Asking for secrets, private data, or medical advice the model should not invent.
Chasing the perfect prompt instead of iterating twice with feedback.
Stacking five unrelated asks so the model averages them into mush.
Treating a fluent paragraph as proof the constraints were followed.

Two-iteration rule

First pass: get structure.
Second pass: paste what is weak and ask for a targeted fix (“shorten bullets 2 and 4; remove any claim without a source”).

If you need a third pass for the same job, you probably need a human source, not another adjective.

Take the improved ask into Verify before anything leaves your machine. If the task is high-risk paste or send, read Verify before you paste. Keep the principles checklist nearby.