Appendix E — A Prompt Library for Writing and Revision

Ready-to-adapt prompts for using an LLM as a writing tool, built on Chapter 29's principles. Every prompt comes with a note on how to evaluate the output — because a prompt you can't grade is a prompt you shouldn't run.

Chapter 29's reframe governs this whole appendix: AI is for revising thinking you've already done, not for replacing the thinking. The model produces plausible text, not true text — most plausible text is correct (that's why it's useful) and some is wrong while looking identical (that's why it's dangerous). So each prompt below assumes you did the thinking and you stay the judge.

The governing rule, before any prompt: If you cannot evaluate whether the output is correct, do not use AI for that task. The tool is safe exactly as far as your judgment reaches — and most dangerous in the very situation where it's most tempting: when you don't know the subject and want the model to know it for you.

A note on the four levers. Good prompts set role, audience, constraints, and examples — which are just Chapters 2, 4, and 7 aimed at a machine. Audience is the highest-leverage lever; you can't specify a reader you haven't thought about, which is how a good prompt sneaks the thinking back in. Fill the [bracketed] slots with real specifics; a vague prompt yields generic mush.

1. Brainstorm — generate options before you commit

Use when you're stuck at a blank page and want raw material to react to, not a finished answer.

I'm writing [document type] for [specific audience] about [topic].
The single thing I want them to do/believe after reading is [goal].
Give me 10 different angles or opening hooks. Be varied — include at
least one I probably wouldn't think of. One line each, no explanation.

How to evaluate: You don't need expertise to judge a brainstorm — you need taste. Throw out the generic ones, keep the two or three that fit your situation, and write the actual draft yourself. The list is fuel, not a draft. Safe even when low-stakes, because you're the judge and nothing ships unchecked.

2. Outline — pressure-test structure

Use after you know your content, to check the order before you draft (Chapter 4).

Here are the points I need to make: [paste your unordered points].
The audience is [who], and they will mostly [scan / read deeply].
Propose a structure that leads with what they most need. Flag any
point that seems out of place or missing. Explain your ordering in
one sentence per section.

How to evaluate: You already know your material, so you can immediately see whether the proposed order serves your reader. Treat it as a colleague's suggestion: adopt the moves that improve flow, reject the ones that don't. Run a reverse outline (Chapter 4) on the result to confirm the skeleton holds.

3. Critique — get a second set of eyes

The model is genuinely good at this, because you supply the substance and it reacts.

Critique the draft below as a skeptical [reviewer type — e.g., busy
executive / peer reviewer / new user]. Where is the main point buried?
Where would this reader get confused, bored, or stop reading? What
claim is unsupported? Don't rewrite it — list specific problems with
line references.

How to evaluate: Each flagged problem is a hypothesis about your draft, not a verdict. Reread the flagged line and decide for yourself whether it's actually unclear — the model sometimes flags fine sentences and misses real ones. Acting on a true critique is augmentation; accepting every note blindly hands away your judgment.

4. Tighten — cut bloat from your own prose

The safest, highest-value use: the model rephrases sentences you wrote (Chapter 3, Appendix A).

Tighten the paragraph below. Cut empty phrases, free buried verbs,
prefer active voice, keep every fact and number. Don't add anything.
Show the result and a one-line note on what you cut and why.

How to evaluate: Diff it against your original. Confirm that no fact, number, hedge, or caveat was dropped or changed — models smooth prose and can quietly delete a qualifier that mattered. You wrote the original, so you know what must survive. If the tightened version says something you didn't mean, keep yours.

5. Translate for an audience — re-aim the same content

Use to adapt a passage you understand for a different reader (Chapter 2, Chapter 28, Chapter 36).

Rewrite the passage below for [new audience — e.g., a non-technical
executive / a patient at an 8th-grade reading level / a domain expert].
Keep the meaning exact. Translate or define any jargon they won't share.
Match their decision: what do they need to DO with this?

How to evaluate: This is where the evaluation rule bites hardest. You can only check the translation if you understand both the content and the target reader. Verify that the simpler version still states every safety-critical caveat (especially in healthcare — Chapter 36) and that the expert version didn't introduce a term you can't defend. If you don't actually know the target audience, you can't grade the output — so don't use it.

6. Mechanical conversion — reformat, don't rethink

Low-risk because the task is purely mechanical and the result is checkable at a glance.

Convert the reference list below from IEEE to APA format. Don't change
any facts — same authors, titles, years. Flag any entry missing
information needed for APA.

How to evaluate: Spot-check several entries against the source and against Appendix C. The model formats wrong metadata flawlessly, so a perfect-looking APA entry can still carry a mistyped year. Mechanical does not mean unverified.

What to never ask AI to do

These map onto the model's "bad at" column (Chapter 29) — original thinking, calibrated nuance, your institutional context, and accuracy on facts you can't check:

  • "Write the whole thing for me" on a topic you don't understand. You'll get fluent text you can't evaluate — the exact failure the governing rule forbids.
  • "Find sources that support this argument." Models fabricate realistic-looking citations for papers that don't exist; lawyers have been sanctioned for filing them (Chapter 11, Chapter 29). Find real sources yourself; use AI only to format the ones you've verified.
  • "Decide how confident I should be." Calibrated hedging requires knowing your evidence (Chapter 7); the model defaults to the average take.
  • Anything requiring your specific situation — your team's history, this client's politics, what your reviewer actually cares about. The model has no access to it.

Integrity checklist (run before anything ships)

  1. Did the idea originate with me? Yes → augmentation, fine. The model supplied the substance → you outsourced the thinking that was your job.
  2. Have I verified every fact, number, and citation? Accuracy is non-delegable. "The AI told me" is not a defense.
  3. Can I evaluate this output? If no, don't use it.
  4. Do I need to disclose? Per your context — academic settings usually require it; many journals now do; workplace usually doesn't. When unsure, disclose.

AI-assisted (you thought, it helped you revise) is usually fine. AI-generated (it supplied the substance and you passed it off) is the problem. The line is the thinking — and the thinking is the whole point of writing (Chapter 1).