Free practical guide
Give Claude real examples of your writing, then get grounded drafts that use your judgment, rhythm, and words without the usual AI tells.
Why this guide exists
Start with one prompt
Paste this into a new Claude conversation. Claude will help choose your samples, learn your voice, build the writing guide, test it on work it has not seen, and improve one real draft. Keep using the same conversation.
Guide me through building a human-writing system from interview to a tested first draft. You own the setup sequence and keep all decisions in this conversation. Do not make me copy another prompt. Use five phases: Interview, Connect, Build, Test, Run. Show the current phase and one decision at the top of each reply. Reuse known answers and keep style evidence separate from factual authority. Explain any Claude feature I may not recognize in one plain sentence. INTERVIEW Ask at most six questions, one at a time. Start by asking for five to seven writing samples I own and genuinely like. Reserve one or two samples as unseen holdouts for testing. Map: 1. My Claude surface and where the system should live: Custom Skill, Cowork folder, Project, or downloadable files. 2. The audiences, formats, decisions, and desired reader actions that matter most. 3. My point of view, directness, warmth, humour, formality, and what must never be softened or exaggerated. 4. Evidence-backed habits in structure, sentence rhythm, vocabulary, punctuation, openings, endings, examples, and calls to action. 5. Phrases, tones, AI patterns, editorial habits, and rewrites that make the work feel less like me. 6. Format-specific differences, factual sources, review boundaries, and how I will judge a strong result. Then show a concise diagnosis with cited sample evidence, five rules to preserve, five patterns to remove, format modes, conflicts between samples, unknowns, and the single most important uncertainty. Ask for one correction. AFTER I APPROVE THE DIAGNOSIS Continue through Connect, Build, Test, and Run in this conversation. Build a source manifest that labels each item Training sample, Holdout sample, Fact source, Template, or Excluded. Never imitate a third party or treat a style sample as factual authority. Build and install the package for my chosen surface. Show the file manifest and meaning-preservation rules before writing or packaging. Run all three tests without using the holdout text as a template. After a failure, repair the system and rerun the case. For the first real draft, show the source map, draft, unsupported claims, and three consequential editorial choices. Ask before publishing, replacing an approved file, or adding a new voice rule. Finish with an approved learning record, not an automatic style change.
What happens next
You answer a few short questions. Claude carries the work through these five stages and pauses when it needs your review.
You are finished when three unseen samples keep your facts and judgment, and one real draft sounds more like you without picking up generic AI habits.
About 25 minutes with five to seven samples you own, then three tests and one real draft.
A clear map of voice, audience, judgment, and patterns to avoid.
Writing samples and fact sources assigned the right role.
Your voice guide, editing rules, good examples, and change record.
Three unseen samples passed after weak rules were fixed.
One real draft checked against its sources and your voice.
Want the details?
Open only the stage you need. These are explanations and recovery points, not five more tasks to complete.
Claude learns what good work looks like for you before it builds anything.
Paste this into a new Claude conversation. Claude will help choose your samples, learn your voice, build the writing guide, test it on work it has not seen, and improve one real draft. Keep using the same conversation.
The interview is concise and stays in one conversation. Claude reuses what you have already said, offers “Not sure” when that is a valid answer, and shows you a short blueprint before making anything.
Claude checks each required action safely and keeps a manual option when access is limited.
No connector is required. Claude should create a small source manifest and use uploaded files first. Connect a document source only when bounded access materially improves calibration, factual grounding, or destination fit.
Resume my human-writing setup at Connect using the approved diagnosis. Create a source manifest before reading more material. Label each item Training sample, Holdout sample, Fact source, Template, or Excluded. Record owner, permission, format, date, audience, authority, and any confidentiality limit. Keep holdout samples hidden from the rules until Test. Guide me through uploaded files first. If a connector would materially help, explain the exact bounded folder or document access needed and never ask for credentials in chat. Test retrieval on one named file. If unavailable, use an upload or paste fallback. Finish with the source manifest, the conflicts or gaps you found, and one correction question before Build.
Claude creates the instructions, examples, templates, and review rules behind the result.
Resume my human-writing setup at Build. Use the approved diagnosis and source manifest. Build for the installation path we approved. For a Custom Skill, create an installable human-writing ZIP with the folder at its root and a valid SKILL.md. For Cowork, write only inside the approved folder. For a Project, provide project instructions and exact knowledge files. Without file access, provide downloadable files and a ZIP. Create a focused package with voice modes by audience and format, evidence citations for each rule, meaning-preservation rules, fact and quote checks, anti-AI editing rules, preferred and rejected patterns, before-and-after examples, unresolved questions, holdout tests, setup instructions, and a change log. Do not include holdout wording in the training rules. Show the manifest, install plan, and the rules most likely to distort meaning first. Ask for approval before writing or packaging. After approval, create the files, validate references, guide me through enabling the package, and run one trigger smoke test. Do not invent facts, quotes, experience, results, or opinions.
human-writing/SKILL.mdValid metadata, triggers, source handling, format selection, editing, verification, and review.
human-writing/references/voice-profile.mdEvidence-backed voice rules, format modes, examples, conflicts, and unknowns.
human-writing/references/anti-ai-rules.mdWeak patterns, diagnostic questions, stronger editorial moves, and meaning safeguards.
human-writing/references/source-manifest.mdTraining, holdout, fact, template, and excluded sources with authority labels.
human-writing/examples/Approved before-and-after examples that show the rule without flattening the writer.
human-writing/tests/calibration.mdHoldout inputs, expected traits, observed results, corrections, and rerun decisions.
human-writing/SETUP.mdInstallation, trigger test, source update, disable, and removal steps.
human-writing/CHANGELOG.mdApproved voice changes tied to evidence and dated calibration results.
Claude tries normal, incomplete, and difficult examples, then fixes what fails.
Resume my human-writing setup at Test. Confirm that the installed skill or instructions load, then reveal the reserved holdout samples only for evaluation. Run the email, proposal, and long-form cases from this guide using new source material. For each test, record: audience, purpose, facts that must survive, expected voice traits, draft, unsupported claims, observed traits, pass or fail, correction, and rerun result. Compare against the holdout evidence without copying its topic, sentences, or distinctive phrasing. Check factual fidelity, judgment, specificity, rhythm, format fit, false warmth, hype, filler, mechanical contrasts, dash-heavy punctuation, slogans, and unfinished fields. Do not use an AI detector as a quality score. After a failure, change the smallest underlying rule that explains it, add an example only when needed, increment the version, and rerun the affected case. Finish only when all three pass without erasing necessary nuance or making every format sound identical.
A concise follow-up with a real decision, one commitment, and a clear next step.
It sounds natural aloud, uses the correct facts, and avoids filler, false warmth, and invented familiarity.
A recommendation that needs evidence, confidence calibration, and a clear commercial point of view.
The argument is specific and supported without hype, fake certainty, or generic benefit language.
A complex topic for a smart non-expert, using approved sources and necessary qualifiers.
The structure is easy to follow, the rhythm varies naturally, and simplification does not erase important nuance.
You use one simple request, review the prepared work, and approve only the actions you want.
Use my installed human-writing system to [draft or edit] this [format] for [audience]. The point is [point], the desired reader action is [action], and the factual sources are [files or links]. Before drafting, show a short source map that separates facts, my judgment, quoted language, and missing information. Choose the relevant voice mode. Preserve necessary qualifiers and mark unsupported claims. Return the finished draft, a source check, the three editorial choices that most affected meaning or voice, and a short run receipt. Do not publish or replace an approved file. Ask before adding any correction from this job to the canonical voice profile, then record only the change I approve.
Do not call the first run complete until
Keep control