The prompt rewrite: turn three weak prompts into keepers (Day 2 of the 30-Day Challenge)

The task

Take three prompts you actually use — an email ask, a research ask, a summary ask — and rebuild each with the six-part anatomy: role, context, task, constraints, examples, format. This is the Day 2 build from the 30-Day AI-Native Challenge, fully worked.

Before AI

You type "write an email to my client about the delay," get something generic and slightly off-tone, edit half of it manually, and conclude AI is overrated. The prompt was the problem.

What you'll need

• Any chat model (ChatGPT, Claude, or Gemini — free tiers are fine)
• Three real prompts you've actually typed recently (pull them from your chat history)
• A doc to save the winners in — this starts your Prompt Library

The workflow

1. Grab the weak version. Real example — the kind everyone types:

write an email to my client about the project delay

2. Rebuild it with the anatomy. Role → context → task → constraints → examples → format:

You are my chief of staff drafting client email in my voice: direct, warm, no corporate filler.

Context: Acme's dashboard project is slipping 2 weeks because their API access
arrived late. Relationship is good; this is the first delay. Client contact: Dana,
VP Ops, hates fluff.

Task: Draft an email telling Dana about the 2-week slip, what caused it, and the
new date (July 18).

Constraints: Under 150 words. No apology spiral — one acknowledgment, then forward
motion. Do not blame their team for the API delay; state it neutrally.

Format: Subject line + body. End with one concrete next step, not "let me know
if you have questions."

3. Run both. Compare. The weak prompt produces a generic apology. The rebuilt one produces something you'd send with one small edit. That delta is the whole skill.

4. Repeat for a research ask. Weak: what should I know about AI regulation. Rebuilt:

You are a policy analyst briefing a startup founder with 5 minutes.

Task: The 3 AI-regulation developments from the last 6 months most likely to
affect a small US SaaS company using LLM APIs.

Constraints: Only things already enacted or formally proposed — no speculation.
Name the jurisdiction and the date for each.

Format: 3 bullets, each: what happened → who it affects → one action to consider.
Flag anything you're not certain about with [VERIFY].

5. Repeat for a summary ask. Weak: summarize this document. Rebuilt:

Task: Summarize the attached document for someone deciding whether to read it.

Format: (1) One-sentence "what this is." (2) The 3 claims that matter, each with
the page or section it came from. (3) One thing the author assumes but never
defends.

Constraint: If you can't find a claim's location, say so instead of guessing.

6. Save all three to your Prompt Library with a one-line "use when." These are your first entries — the library is challenge artifact #2, and it compounds all month.

Verify it worked

Run the weak and strong versions of the same ask side by side. The strong output should need one edit or fewer to actually use. If it needs three, your context or constraints section is too thin — add the detail you'd give a new hire.

Troubleshooting

• Output still generic? Your context block is describing the task, not the situation. Add the names, dates, stakes, and tone notes you'd give a human.
• Too long / overwrought? Constraints beat instructions. "Under 150 words" outperforms "be concise."
• Model ignores the format? Put format last and make it mechanical ("Subject line + body", "3 bullets, each: X → Y → Z").

Reality check

The anatomy isn't magic — it's just the information a competent assistant would have needed anyway. If you can't fill in the context block, the model can't either; that's a signal the task isn't ready to delegate.

Data & security

Real client names and project details are fine in a tool your org has cleared; on a consumer tier, sanitize them (see Day 6 of the challenge — the Data-Boundary Playbook).

Going further

Day 3 of the challenge runs one of these prompts across models and tiers to build your Model & Cost Scorecard. And once you have five keepers, package the best ones as reusable skills — the SKILL.md workflow shows how.

Your takeaway

Three rebuilt prompts, saved with "use when" notes — the first entries in a library that makes every future session faster than the last.