Five parts that hold up.
Strong prompts have five parts. A role. A clear goal. Constraints, like length or tone. Examples of what good looks like. And the exact output format you want back.
prompt.md
Role: Senior product marketer.
Goal: Launch copy for the pricing page.
Constraints: Under 120 words. No hype.
Example: see docs/voice.md
Output: 3 headlines + 1 paragraph.
- Role — Role — who it should be
- Goal — Goal — what success is
- Limits — Constraints — the rules
- Example — Examples — show, don't tell
- Output — Output — exactly what to return
Write it like a brief for a smart colleague.
Point to files instead of pasting them in.
Put together, it reads like a brief you would hand a smart colleague. Notice the example points to a file, instead of pasting it in.
Iterate. Don't restart.
When the output misses, resist the urge to start over. Say exactly what's wrong, and what to keep. Starting fresh throws away everything the agent just learned.
Start a new chat vs Steer the same thread
| Tempting | Do this |
|---|---|
| Loses everything it learned | "Para 2 is too salesy — cut it." |
| Same mistakes, round two | "Keep the headlines as they are." |
| You re-explain from zero | Context and progress stay intact |
Plan first. Then build.
For bigger jobs, use a loop. Ask for a plan. Correct the plan. Execute it, one step at a time. Then verify the result.
- Plan — ""
- Correct — ""
- Execute — ""
- Verify — ""
What you just learned.
- Role, goal, constraints, examples, output
- Iterate in the same thread — don't restart
- Plan → correct → execute → verify
Your mission: Rewrite one weekly prompt with all five parts — then run the plan-first loop.