Chapter BAI for writingPage 8 of 8

AI for writing

Mastery: build your writing prompt pack

A useful prompt pack turns writing standards into repeatable checkpoints while keeping evidence, judgment, and final authorship human.

~17 minMastery check

Before you start

Why this matters

Name three writing tasks you repeat: perhaps a proposal, explainer, report, lesson, or update. For each, write the failure that matters most. Is it an unclear audience, generic structure, unsupported claim, lost qualification, inconsistent voice, or incomplete review? Your prompt pack should prevent real failures, not collect impressive wording.

1Learn the idea

Read

Build a workflow, not a magic prompt

One enormous prompt asking for research, planning, drafting, editing, style, and verification makes changes difficult to inspect. Use small templates with one job and a human gate between stages.

Every template should identify:

  1. purpose and intended user;
  2. required inputs;
  3. allowed transformation;
  4. facts or meaning to preserve;
  5. prohibited additions or changes;
  6. uncertainty behavior;
  7. review output;
  8. accountable human decision.

Store visible placeholders. A blank [PRIMARY AUDIENCE] should not disappear into generated assumptions. Add a “when not to use” note, especially for restricted data, prohibited academic assistance, high-stakes advice, or unavailable sources.

Read

Prompt 1: brief before prose

Help me create a writing brief. Do not outline or draft yet.

Document type:
Purpose: after reading, the audience should...
Primary audience and context:
What the audience already knows:
Core message:
Approved evidence and sources:
Scope and exclusions:
Format and constraints:
Success criteria:
Permissions, integrity, or sensitivity rules:

Identify missing information, contradictions, unsupported ambitions, and
audience misunderstandings. Ask up to seven questions. Preserve unknowns
as [NEEDS INPUT] rather than guessing.

Use this template whenever the assignment is unclear. The human approves the brief before structure work begins.

Read

Prompt 2: compare outlines

Using the approved brief, propose three genuinely different outlines.
For each, provide:
- organizing logic and reader journey;
- section titles and section promises;
- reader question answered by each section;
- required evidence;
- transition to the next section;
- approximate word budget;
- strengths, risks, and omissions.

Do not draft prose, invent facts, or add sources. Flag any section whose
promise cannot be supported by the approved evidence.

Choose a structure based on audience and purpose, not novelty. The human freezes a working outline and records later structural changes.

Read

Prompt 3: bounded section draft

Draft only the named section.

Section promise:
Reader context:
Approved evidence:
Required terms and qualifications:
Continuity notes:
Word budget:

Use only supplied evidence. Do not invent quotations, examples, statistics,
sources, causes, consensus, or outcomes. Mark missing support [SOURCE NEEDED].
Return two meaningfully different versions and then list every factual claim,
certainty phrase, and assumption in each.

Drafting alternatives should test a real choice, such as direct versus scenario-led explanation. The claim list is a review aid, not verification.

Read

Prompt 4: clarity diagnosis and edit

First diagnose without rewriting:
- section and paragraph jobs;
- repeated ideas;
- missing reasoning;
- buried points;
- unclear actors, references, terms, or transitions.

Then propose edits that preserve facts, logic, uncertainty, conditions,
numbers, names, quotations, terminology, attribution, and requested action.
Do not add evidence, strengthen certainty, infer motives, or change authority.
Return a change log and flag every meaning-sensitive edit.

Apply structural changes before sentence polishing. Compare the source and revision yourself.

Read

Prompt 5: voice and style pass

Apply the approved voice card to this clear draft.

Stable voice rules:
Context-specific tone:
Preferred terminology:
Anti-patterns:

Preserve all verified meaning, qualifications, citations, quotations,
technical terms, and authority boundaries. Do not add anecdotes, emotion,
praise, urgency, or opinion. List the style choices applied and flag places
where style conflicts with precision.

Do not use this prompt to imitate a living writer or misrepresent authorship. Use permitted examples and high-level features.

Read

Prompt 6: claim ledger

Do not rewrite and do not claim to verify.
Extract every factual claim, name, number, date, quotation, attribution,
comparison, causal statement, externally testable recommendation, and link.

Return:
- exact claim and location;
- claim type;
- source needed;
- scope, time, denominator, and certainty to inspect;
- whether the wording appears stronger than the supplied source notes.

Mark source status UNKNOWN until a human opens an authoritative source.

The human completes the ledger using inspected sources. Generated citations and search snippets do not count as verification.

Read

Prompt 7: revision and final review

Against the approved brief, outline, voice card, and completed claim ledger,
identify the five highest-impact revision needs. Prioritize purpose, logic,
support, reader action, and harmful ambiguity over cosmetic preferences.

For each, show a short excerpt, explain the problem, propose a bounded change,
and list checks that must be repeated. Do not silently edit or approve.

End with an OPEN ITEMS checklist covering placeholders, unsupported claims,
permissions, links, accessibility, required disclosure, qualified review,
and accountable final approval.

The final decision belongs to the author or designated reviewer. Never treat “the model found no issues” as approval.

Read

Test and maintain the pack

Test prompts with normal and adversarial cases:

  • missing audience and contradictory goals;
  • a compelling outline with no supporting evidence;
  • invented quotations or plausible statistics;
  • a rewrite that removes “may,” “only,” or “unless”;
  • a style request that adds false enthusiasm;
  • confidential material in an unapproved tool;
  • a high-stakes claim needing qualified review;
  • an assignment where AI drafting is prohibited.

Define observable pass criteria. Did unknowns remain visible? Did the outline expose evidence gaps? Did drafting preserve denominators? Did editing retain conditions? Did the claim ledger find causal language? Severe errors matter more than average fluency.

Version the pack, name an owner, and review it after policy changes, tool changes, incidents, or repeated failures. Remove templates people no longer understand. A prompt is operational only when another person can identify required inputs, prohibited uses, and the human gate without your explanation.

Read

Demonstrate mastery

Complete one document from brief to final proof. Submit the approved brief, three outline alternatives and your rationale, two bounded section drafts, a continuity sheet, a preservation edit, a voice card, a human-completed claim ledger, a revision log, and the final approval checklist.

Explain one point where you rejected generated prose, one where you narrowed a claim, and one where you stopped because evidence or permission was missing. Mastery is not maximum generation. It is controlled assistance with defensible authorship.

Checking tutor…

Continue learning · glossary & guides
  • Why should writing stages use separate prompts?
  • What eight elements belong in every reusable template?
  • Which outputs can assist verification but never replace it?
  • How should a prompt behave when evidence or permission is missing?
  • What proves a prompt pack works for someone besides its creator?
  • Cheatsheet: prompt recipe · Glossary: responsible AI