Claude writing workflow: 10 prompts that replace a content agency

Why agency content misses

Author’s claim: an agency delivered three articles a week, structurally correct and on topic, with none of the writer’s voice. Editing took longer than writing from scratch, and engagement ran at about half the writer’s own posts.

Root cause: a brief describes a voice in terms a third party can’t reproduce. Fix: a configured workspace that carries the voice itself: never-list, audience description, example posts, product facts. Below: the setup, then the 10-prompt pipeline, upgraded.

Build a voice file from evidence

“Professional but friendly” steers nothing. A voice file does: tone (and what you are not), sentence rhythm, how you open and close, words you use, words you never use, formatting habits, and the mistakes AI drafts make that you’d correct.

Add 2–3 before/after pairs: a generic AI line next to your real one. A pair shows the transformation; a rule only describes it. The never-list carries the most weight, because your taste shows in what you reject.

Ready-made prompt and template: see the Snippets “Voice file builder” and “Voice file template” below. Pick 4–6 samples you’re proud of, across formats (long article, email, social post), not just your latest.

Split the work: Project vs Skill

Per Composio’s Aug 2026 guide: a Claude Project loads its instructions and files into every chat inside that workspace. A Skill is a folder of instructions Claude loads on demand, and it works inside and outside Projects. Rule of thumb: the Project holds what Claude needs to know, the Skill holds how to do a repeatable job.

Custom Styles: a June 2026 guide reports Anthropic’s docs describing Styles as moving to Skills, with both still working. If you rely on a Style for voice, check Settings and plan the move.

Keep the voice in one place

A Sep 2026 workshop breakdown from Zenler flags that AI memory settings can interfere with consistent voice training. Work inside one Project that holds the voice file, so the voice has a single source. Test: run the same angle prompt inside and outside the Project and compare what changes.

1. Angle generator

The blank page is an angle problem, not a writing problem. Once the angle is set, sentences come fast. Upgrade: you feed in something only you have (a story, number or opinion), and generic angles are dropped up front.

I'm writing about [TOPIC] for [SPECIFIC READER].
What I bring: [ONE REAL STORY, NUMBER OR OPINION I CAN BACK UP].

Give me 7 angles. Each is one sentence that takes an arguable position, not a description of the topic. Each should state something this reader hasn't seen put that directly. Drop any angle a generic article on this topic would already make. Numbered list, no explanations.

2. Argument anchor

Seven angles isn’t a decision. This forces one. Even when you disagree with the pick, a concrete alternative sharpens what you want to argue. Upgrade: the pick must be carried by your evidence, and the model names the runner-up plus the risk.

Here are the angles: [PASTE LIST].
My reader: [SPECIFIC SITUATION OF TARGET READER].
Evidence I have: [WHAT I CAN PROVE OR TELL FIRST-HAND].

Pick the angle that gives this reader the most specific, usable piece and that my evidence can carry. Give two sentences on why it wins. Then name the runner-up and the one risk that could sink your pick.

3. Outline builder

“Every section must move the argument forward” kills filler sections. Two additions: sections without evidence get flagged instead of padded with invented support, and a cut test exposes the weakest section before you write it.

Build an outline for a [WORD COUNT] article on this angle: [ANGLE].
Structure: a hook section, 3-4 body sections that each make one distinct claim advancing the argument, and a closing section stating the implication.
For each section give: the heading, the claim in one sentence, and the evidence or story that would make it specific. Where I haven't given you evidence, mark the section "NEEDS EVIDENCE" instead of inventing any.
Then name the one section I could cut with the least loss to the argument.
My material: [NOTES, STORIES, NUMBERS].

4. Section writer

One section at a time beats a full draft for consistency. The closing-transition rule does the most work: it stops sections from ending on a recap that repeats the point. Upgrade: hard limits like “max two sentences” and “no em dashes” now live in the voice file, and rhythm variation is requested so the result doesn’t read as a staccato list.

Write one body section. Heading: [HEADING].
It argues: [ONE SENTENCE].
Evidence or story to use: [DESCRIPTION]. Use only what I've given you; don't invent numbers, names or quotes.
Follow my voice file. First person, active voice, short paragraphs of one to three sentences. Vary sentence length so it doesn't read as a list of fragments.
End with one sentence that opens the next question without summarizing what you just said.

5. Hook writer

The opening line carries more weight than any other, and a weak one sinks everything after it. Five options across different approaches beat committing to the first idea. Upgrade: hooks may only use facts you supply, so a “number” hook can’t be made up.

Write 5 opening hooks for an article with this angle: [ANGLE].
Reader: [SPECIFIC READER DESCRIPTION].
Mix the approaches: open on a moment, a number, or a contrarian claim. Use only facts I've supplied: [FACTS]. No setup sentences that announce what the piece is about. End each hook on a line that pulls the reader forward.
Numbered list only.

6. Voice pass

Run a systematic voice pass before you edit, so the edit is about quality, not cleanup. Upgrade: it works from the voice file, protects the flow by asking for varied rhythm, and reports any sentence where the change might have shifted meaning, so you check those instead of rereading everything.

Rewrite this draft in my voice using the voice file. Keep the meaning, every core point and every fact; don't add or remove information. Change only how it sounds.
Vary sentence length where the rhythm goes flat. Remove anything on my never-list: [PASTE LIST].
Return the rewrite first. Then list any sentence where your change could have shifted the meaning.
Draft: [PASTE DRAFT]

7. CTA builder

CTAs written cold describe the product from scratch, as if the reader hadn’t just read something relevant. Tying the CTA to the exact topic covered makes it feel earned. Upgrade: no invented results or testimonials. Also fixed: the source prompt hard-coded a profile URL inside the link format; it’s now a [URL] placeholder, so point it at your real destination. Swap to an HTML link if you publish on WordPress.

Write a CTA for the end of this article.
Product: [NAME] - [ONE SENTENCE ON WHAT IT DOES].
Article topic: [ONE SENTENCE].
It should read like a recommendation from someone who uses the product. Tie it to the exact problem this article just covered; don't describe the product from scratch. Claim only what the product actually does: no invented results, testimonials or numbers. No urgency language, no exclamation points, 4 sentences maximum.
End with one link: [ANCHOR TEXT]([URL]).

10. Revision prompt

“Improve this piece” triggers global edits you didn’t want. Three named weak points with rewrites keep the revision contained and the piece yours. Upgrade: two rewrite options per point, plus a list of unsupported claims that flags problems without touching them.

Read this piece and find its three weakest points: the opening line, the section with the least evidence, and the closing sentence.
For each: one sentence on what is weak, then two rewrite options.
Then list any claim that has no evidence behind it. Flag these; don't rewrite them.
Touch nothing else.
Piece: [PASTE]

8. Email adapter

An email is shorter, more direct and ends on a harder CTA than the article, so it’s a translation, not a compression. The length cap stops it from becoming a squeezed copy. Upgrade: subject line options and a preview line, and a rule that the email works for someone who never opens the article.

Turn this article into an email of about 200 words.
Give: 3 subject line options, one preview line, one hook sentence, one paragraph that delivers the core argument, and one CTA sentence pointing to [URL OR PRODUCT].
First person throughout. Start with the hook; no summary of what the email contains. The email must work for someone who never opens the article.
Article: [PASTE]

9. Threads repurposer

Repurposing without a format constraint fits no platform. This encodes the Threads shape: standalone opener, one point per post, action close. Upgrade: every post must also make sense when seen alone, since feeds break threads apart.

Turn this article into a Threads thread of 5-7 posts.
Post 1: the strongest single claim in the article, as a standalone statement that works without the rest.
Each following post develops one supporting point and makes sense on its own if seen alone. The last post ends with one clear action.
2-3 sentences per post. No hashtags. First person.
Article: [PASTE]

Critique before you rewrite

When a draft drifts, “try again” gets another generic draft. Make Claude audit itself against the voice file first, then fix only the flagged sections. Composio’s Aug 2026 guide recommends this loop.

Review the draft above against my voice file. List the 3 places where tone or structure breaks my rules, quoting the exact phrase for each. Then rewrite only those sections and leave the rest intact.

Fix the file, not the draft

Every recurring miss is a missing rule. Add the offending phrase to the never-list, or add a before/after pair that shows the fix. The next draft starts better; a patched draft only fixes itself.

Keep instructions modular

Composio’s guide warns that very long instruction sets make the model weight rules unpredictably, and contradictions compound. Small focused units (outline, headline, edit pass) are easier to update: when one produces bad output, you know which block to revise. The 10 prompts above are that modularity in practice.

Carry decisions across sessions

Style drifts over days or weeks, not within one chat. Keep a running state document in the Project: angle chosen, structure decisions, formatting choices, phrases already used. Paste the relevant part at the start of each new session so Claude doesn’t reset your choices.

Add the human pass

AI executes patterns; judgment and lived detail stay yours. Read the final draft out loud, rewrite any paragraph that performs expertise instead of sharing it, and add one detail only you could know. That single detail is what makes a draft read as written by you.