Content production

How the pipeline works

Every fact comes from the client. Every sentence sounds like the client because it is built from recordings of the client talking. A person decides the topic, runs the interview, and signs off the final piece. Claude does the collating, the structuring, the first draft and the checking.

Person does this Claude does this Stops until approved

Set up / once per client

Client foundations

Then kept topped up

Four files that get built once and reused on every post afterwards. This is the part that makes post twenty better and cheaper than post one. Skip it and every piece starts from nothing.

Person
Choose the voice sources. Recordings of the client actually talking, unscripted.Deliberately not the website copy, and not scripts we wrote for them to read. Both teach the wrong voice.
recordings
Claude
Build the voice profile: sentence rhythm, vocabulary, formality, first person or third, and the words they never use.
voice-profile
Claude
Build the facts bank: pricing, service areas, accreditations, guarantees, and the stories they retell.
facts-bank
Person
Write the no-go list: claims we cannot make, competitors we cannot name, terminology to avoid.A person has to write this one. There is no source document for it.
no-go
Claude
Sweep Drive for new meeting transcripts, match them to the client, and file them.Pull the full transcript rather than the auto summary. Summaries strip out exactly the specifics we need.
transcripts/
Approval

Account manager reads the voice profile and facts bank and corrects anything wrong before a single post is written.

Phase 1 / per post

Research

Work out what people actually want to know about this topic, and what the pages already ranking have failed to answer.

Person
Pick the topic and the target keyword.
brief
Claude
Pull search data, People Also Ask questions, and read the competing pages to find the gaps.
research.md

Phase 2 / per post

Skeleton

The headline plus every subheading, written as the real questions a customer would ask. This doubles as the interview script.

Claude
Turn the research into an H1 and a set of H2s phrased as questions.
skeleton.md
Approval

A writer approves the skeleton before it goes anywhere near the client. A bad skeleton wastes the client's time on the call, and the call is the expensive part.

Phase 3 / per post

Interview

Human led

This is the part that makes the finished piece worth reading, and it is done entirely by a person.

Person
Send the questions ahead, then run a recorded call of roughly twenty minutes.Written answers give you two sentences of nothing. A call gives you the anecdote, the objection they hear every week, and the throwaway line that becomes the hook.
interview.md
Person
File the raw transcript. Do not summarise it first.
transcripts/

Phase 4 / per post

Draft and self-check

The only automated loop

Claude writes a draft using all six files, then checks its own work against three tests that a person defined. If any test fails it rewrites and tries again. Nothing reaches a human until all three pass.

The loop / rewrites until it passes or hits the cap
Write
draft
Test 1VoiceCompare against recordings of the client talking. Flag anything that does not sound like them.
Test 2SubstanceEvery claim traced back to the interview, the facts bank or a cited source. No filler paragraph that could sit on a competitor's site.
Test 3DetectorScore below our threshold on an AI detector. A floor to clear, never the target.
All three
pass →
↻ Any test fails, revise and run againStops after 6 attempts and asks a person for help

Why three tests and not just the detector: a detector only measures sentence variety and word predictability. Chase that score alone and the writing gets deliberately lumpy to game the tool, reading oddly to a human while technically passing. The voice and substance tests are what stop that happening.

Claude
Hand over the draft with a note on which tests passed and on which attempt.
draft.md

Phase 5 / per post

Edit and sign-off

Human led

A draft that passes three automated tests is still a draft.

Person
A writer edits properly: cuts, sharpens, adds what the machine could not know.
draft.md
Sign-off

The client approves before anything is published. Same as it works today.

Person
Capture what the writer and the client changed, and why.This is the raw material for the next step, so it needs recording rather than discarding.
edits.md

Phase 6 / per post

Publish and monitor

Then ongoing

Person
Publish.
live
Claude
Track rankings and traffic, and report what is working.
report

↻ The flywheel

What makes it improve

The setup files are not static. Every edit and every client call feeds back into them, so the system gets more accurate the longer a client is with us. This is the bit that is hard to copy, and the reason it is worth doing properly for one client first.

Writer and client edits sharpen the voice profileedits → voice-profile
New call transcripts top up the factstranscripts/ → facts-bank
Anything the client vetoes joins the no-go listedits → no-go

What the machine never does

Hard limits, built into the process
  • Invent a fact, a price or a claim. Anything it cannot trace fails test two.
  • Decide what is true. That is what the interview and the facts bank are for.
  • Speak for the client using words the client never said.
  • Sign anything off.
  • Publish. A person does that, every time.
  • Learn the client's voice from generic web copy, or from scripts we wrote for them.

What the team gets out of it

The honest version
  • The hours currently lost to collating search data and formatting headings go back to writers.
  • The specifics clients mention on calls stop evaporating, because they are captured in a file instead of one person's memory.
  • Writers spend their time on the interview and the edit, which are the parts that need a person.
  • A new starter inherits the client's voice as a document rather than learning it over six months.
  • Post twenty is better than post one, because the assets improve every cycle.