Asking Claude to write a document
Voice, structure, source-pinning — the difference between a brief and a fight.
Why prose is harder than design
A bad design render, you can see is wrong in two seconds. A bad prose render reads fine. The sentences are grammatical. The paragraphs flow. You can’t tell at a glance that the voice is wrong, or that two of the “facts” were invented, or that the structure quietly buried the actual point.
Prose hides its failures better than design does. Which means the discipline has to be tighter. Three pins, every time: voice, sources, structure. Without all three, the default-LLM voice produces competent-sounding prose that isn’t yours and isn’t reliable.
These three pins sit on top of the five-field brief from Lesson 04. The five fields define what the artifact is for; the three pins define what it sounds like, what it’s based on, and how it’s organized.
Pinning your voice
Voice match works better than most operators expect. The agent doesn’t need a style guide — it needs a sample. 200-400 words of your existing writing, pasted directly into the brief, with “match this voice.”
The trick: have the agent reflect the voice signatures back before drafting. Three to five specific patterns, each falsifiable. “Short paragraphs.” “Bolded one-liners as transitions.” “Ending sentences that turn rather than summarize.” If the signatures match how you actually write, the draft will land. If they don’t, you correct before render — which is much cheaper than correcting after.
Voice match degrades when:
- The sample is too short (under ~150 words — not enough signal)
- The sample is in a different register than the target (a casual essay sample for a formal report won’t transfer)
- You ask the agent to match voice on something heavily structured (legal docs, code comments — the structure overrides voice)
Three ways AI writing fails
The patterns that produce the “reads fine but isn’t mine and might be wrong” output. Hover any card to see the diagnosis.
The default-LLM voice
The unsourced claim
The structureless dump
The fix in all three: voice sample, source list, structure outline — before any rendering. The pins are the gate. Without them you get average prose; with them you get yours.
The voice-sample test
The diagnostic for voice. Paste your sample, get back three falsifiable voice signatures, confirm before drafting.
I'm about to ask you to write something in my voice.
Before drafting, do this:
1. I'll paste 200-400 words of my own writing.
2. You return three "voice signatures" — specific patterns
in my writing, each falsifiable. Examples:
- "Short paragraphs (1-3 sentences average)"
- "Bolded one-liners as section transitions"
- "Ending sentences that turn rather than summarize"
3. I'll confirm or correct the signatures.
4. Only then do you draft, matching the confirmed signatures.
The point: your voice match is more reliable when we
agree on what my voice actually IS before you start.Forces the voice match to be concrete and confirmable before any draft begins.
After you’ve done this 5-10 times you’ll know your own voice signatures by heart. From then on you can paste them directly without the sample-and-extract round. But early on, do the round — it’s how you learn what you sound like.
Annotated transcript
A real conversation: operator asks for a README, agent forces the three pins, agent self-grades the draft on voice match. Click any pinned annotation to expand the “why this turn matters” commentary.
Three operator turns, three agent turns. Result: a draft that sounds like the operator and only uses the facts the operator gave.
- operator
I need to write the README for a tool I'm publishing. Can you draft it?
- agent
Before I draft, three pins:
1. Voice — paste 200-400 words of your existing writing and I'll match it. Without a sample I'll default to standard README-voice, which reads competent but generic.
2. Sources — what's the tool, what does it do, who has used it? I'll only make claims you've given me — no inventing usage stats or testimonials.
3. Structure — standard README order is fine (what → why → install → use → license) unless you want something different. Confirm or override.After the three pins, one render, then we revise the BRIEF if it doesn't land.
- operator
Voice: [pastes 300 words from a recent essay]. Sources: tool is the brief format from Lesson 04, used across all my own work for ~8 weeks, ~30 briefs run, no public usage yet. Structure: standard README plus a "why this exists" section before install.
- agent
Two checks before I draft:
- "~30 briefs run" is a real number. I'll use it. "No public usage yet" I'll either skip or surface as "adopted internally; published today" — your call.
- Your voice sample has three signatures: short paragraphs, occasional bolded one-liners, and ending sentences that turn rather than summarize. I'll match those.Flag if either is wrong before I draft.
- operator
Surface as 'adopted internally; published today.' Voice signatures match. Draft.
- agent
[draft delivered — 600 words, three short paragraphs in the opener, a bolded one-liner mid-page, ending with a turn rather than a summary]
My own pre-test read: voice match is ~85% — I caught the short paragraphs and the bolded line, missed your habit of starting paragraphs with subordinate clauses (yours do this 4 times in the sample, my draft only twice). Sources: only "~30 briefs / 8 weeks / adopted internally" used; no invented numbers. If you want stronger voice match, I can revise pass-2 specifically targeting subordinate-clause openings. Otherwise ship.
Writing brief template — copy and paste
Save this in a snippets file. Paste before any prose ask. Fill all three pins; mark FUZZY for any you haven’t resolved.
Brief: <one-line title of the artifact>
[Five fields from Lesson 04: who, what, why, where, how to test]
VOICE PIN
---------
Sample (200-400 words of my writing):
<paste here>
Match instruction:
- Extract 3-5 voice signatures and reflect them back
- Wait for confirmation before drafting
SOURCE PIN
----------
The ONLY facts you may use in the draft:
- <fact / number / quote — with attribution if applicable>
- <fact / number / quote — with attribution if applicable>
- <fact / number / quote — with attribution if applicable>
Rules:
- Every factual claim must trace to a source above
- If a claim is helpful but unsourced, hedge or remove
- No invented numbers, testimonials, or attributions
STRUCTURE PIN
-------------
Section order, one phrase each:
1. <section 1>
2. <section 2>
3. <section 3>
...
After draft, self-grade on voice match (with specific
signatures hit/missed) and source fidelity (any
unsourced claims listed).
The self-grade at the end is non-negotiable. Without it the draft reads fine and you don’t know what to fix. With it, the agent points at its own weak spots and you decide whether they matter.
Prompt kit
Three prompts for voice-pinning, source-pinning, and grading. Save in your CLAUDE.md or a personal snippets file.
I'm about to ask you to write something in my voice.
Before drafting, do this:
1. I'll paste 200-400 words of my own writing.
2. You return three "voice signatures" — specific patterns
in my writing, each falsifiable. Examples:
- "Short paragraphs (1-3 sentences average)"
- "Bolded one-liners as section transitions"
- "Ending sentences that turn rather than summarize"
3. I'll confirm or correct the signatures.
4. Only then do you draft, matching the confirmed signatures.
The point: your voice match is more reliable when we
agree on what my voice actually IS before you start.I want you to write <artifact>.
Voice: <paste 200-400 words of your own writing>
Sources (the ONLY facts you may use):
- <fact / number / quote with attribution>
- <fact / number / quote with attribution>
- <fact / number / quote with attribution>
Rules:
- Every factual claim must trace to a source above
- If a claim is helpful but unsourced, hedge it
("likely," "in our experience," "anecdotally")
or remove it
- No invented numbers, no invented testimonials,
no invented frameworks attributed to real people
Structure: <list the sections you want, in order>
Before drafting, restate the voice signatures, sources,
and structure back to me so I can correct before render.Here's the brief: <paste voice + sources + structure>
Here's the draft: <paste draft>
Grade on three axes:
- Voice match: 1-5, with the specific signatures hit/missed
- Source fidelity: are any claims unsourced or extrapolated
beyond what I provided? List each.
- Structure: did the draft follow the section order I gave?
For any axis below 4, propose the smallest revision that
would lift it. Don't rewrite — surgical only.Apply this — your next document
20-minute exercise. One real document. Three pins. One render. Self-grade.
Your first three-pinned document
Each step takes 3-5 minutes. Progress saves automatically.
- 01Pick a piece of writing you've been putting off. README, about section, recap, anything.Don't pick the most important piece — pick something with stakes low enough that you'll actually finish the exercise.
- 02Paste 200-400 words of your existing writing into the agent. Run the voice-sample test.Confirm the three voice signatures it returns. Correct any that are wrong before drafting begins.
- 03Pin sources. List every fact, number, or quote the agent is allowed to use. Nothing else is permitted.If your source list feels short, that's fine — the draft will be shorter and tighter than you expected.
- 04Brief the structure. 3-5 sections, in order, one phrase each.If you can't say what the third section is, your structure isn't ready — figure that out before render.
- 05Render once. Read the agent's self-grade. Revise the brief if needed; only revise prose if the brief was right.This is the discipline that breaks the "23 versions" loop in writing the same way it does in design.