Workflow · 11 min
How to build a repeatable AI editorial process for a small team
Most teams using AI for content end up with inconsistent output because they have no process — only individual habits. A documented editorial process is the difference between AI as a multiplier and AI as a quality lottery.
- 1
Step one — define the editorial bar in writing
Before any process, agree on what 'good' looks like. Three concrete attributes the work must always have (e.g. specific examples, clear thesis, named author) and three the work must never have (e.g. generic openers, unverified claims, hidden AI use). Write them down. Pin them to the workspace. Every later step refers back to this document.
- 2
Step two — write a voice document
A two-page document covering: who you write for, what your voice sounds like in three named attributes, what your voice avoids, three worked before/after examples, and house style decisions (Oxford comma, headline case, hedging norms). Most teams skip this and pay for it in inconsistent output forever after. The voice document is the single highest-ROI artefact in the process.
- 3
Step three — build a shared system prompt
Convert the voice document and editorial bar into a system prompt all team members use when drafting with AI. Include the voice attributes, the things to avoid, a worked example, and a short list of generic words the model should not use. Version the prompt. When the voice document updates, the prompt updates with it.
- 4
Step four — separate drafting from editing roles
Even when one person does both, separate the modes in time. Draft with AI in one session; close the tool. Edit in a different session, ideally a day later. The hybrid mode — drafting and editing simultaneously — produces worse work than either alone because it tempts the editor to accept rather than question.
- 5
Step five — run the standard editing pass
Document a single editing checklist every piece runs through: opening (cut throat-clearing), structure (reorder for development), specificity (replace one generic noun per paragraph), hedging (cut by half), examples (one load-bearing per section), ending (replace summary with commitment). The checklist becomes muscle memory and the quality variance drops.
- 6
Step six — verify every factual claim
Build a verification step into the process, not the writer's discretion. A simple shared template: list every factual claim, mark verified/corrected/removed, link to the source. No piece publishes until this list is clean. The discipline takes time the first few times and becomes routine.
- 7
Step seven — second-set-of-eyes review
Every piece longer than 500 words gets reviewed by a teammate, even briefly. The reviewer checks against the editorial bar, the voice document, and the verification list — not against their own taste. Reviews are written as 'meets bar' or 'specific items to fix,' never as discretionary edits. The discipline keeps reviews fast and decisive.
- 8
Step eight — disclose AI use clearly
Adopt a default disclosure line: a one-sentence note near the byline naming how AI was used and who the human editor was. Specific disclosure builds trust. Vague boilerplate does not. The disclosure language itself becomes part of the brand voice.
- 9
Step nine — measure something
Pick two metrics — one input (number of pieces shipped per week) and one output (engagement, reply rate, citation in AI search, whatever matters for your use case). Track them weekly. Without measurement the process becomes ritual; with measurement it becomes craft.
- 10
Step ten — review the process monthly
Once a month, review the editorial bar, the voice document, and the system prompt against the work that shipped. What patterns are recurring? Which checklist items are being skipped? Which examples in the voice document are no longer representative? A process that updates itself stays useful; a process that calcifies stops being followed.
Common pitfalls
- •Adopting a process and then quietly skipping it for urgent work. The urgent work is exactly when the bar matters.
- •Letting individual writers each maintain their own prompt and voice. Inconsistency follows.
- •Treating the process as a constraint to defend rather than a tool to improve. The process is yours; update it.
- •Confusing volume metrics with quality metrics. Shipping more bad pieces is not progress.
- •Skipping verification because the model 'usually gets it right.' The piece you skip on is the one that publishes the error.
Frequently asked
How big does the team need to be for this to be worth it?
Two writers and up. Solo writers benefit from the editorial bar and voice document; the review step requires at least one teammate.
How long does the full process take per piece?
For a 1,500-word article: roughly 45–90 minutes of human time across drafting, editing, verification, and review. The variance is mostly in verification load.
What tools do I need beyond an AI writing tool?
A shared document for the voice and editorial bar, a checklist for editing, a verification template, and a measurement spreadsheet. No specialised tooling required.
Run the workflow on a real draft
Open the rewrite engine and apply the steps to writing you actually have to ship today.
Start rewriting