A writing working agreement: draft, and do not publish
What an agent may write unattended, where it must stop, and the two failures that are specific to prose — inventing detail to finish a sentence, and not knowing what you have already said.
- Authors
- Leon Mallett, Founder of Captivated Ltd with Claude Code
- Status
- Last confirmed working 30 August 2026 on Claude Code 2026-08
- Written
- 30 August 2026
- Licence
- Handover-1.0
Customise before downloading
This document has 3 fill-in fields. Complete them and you get a copy ready to commit, rather than one to edit afterwards. Nothing is sent anywhere — the file is assembled in your browser.
The second half matters more than the first. An agent that does not know what the audience already knows will explain the wrong things at length.
Name the gate. Without it "publish" is undefined and a saved draft in the CMS looks like the end of the job.
The single least inferable thing here, and the most expensive to get wrong. Give the reason as well as the rule — a rule without its reason gets argued around by the next agent.
A customised copy is a derivative of the published document: its contents differ, so it carries a line naming the original. Any field you leave blank keeps its{{placeholder}}, so it stays obvious rather than silently empty.
This document sits at the root of a writing project and tells an agent where its work stops. A folder is enough; nothing here assumes version control.
Who reads this: {{audience}} Publishing means: {{publishes_via}} Never written about here: {{never_covered}}
The division
Drafting is work. Publishing is a decision, and it is irreversible in a way almost nothing in software is — a sent newsletter cannot be unsent, and a correction is a second public act rather than an edit.
So the line sits at the moment the words become someone else’s problem. Write, restructure, cut, rewrite, argue with the brief. Do not publish, send, post or schedule.
Safe without asking
- Drafting, restructuring, cutting, rewriting to length.
- Producing alternative openings, headlines and standfirsts for comparison.
- Checking claims against sources and flagging the ones that do not hold.
- Auditing a draft against the editorial rules and reporting where it breaks them.
- Searching the archive for whether this ground is already covered.
- Saying the brief is wrong.
Stop and ask first
- Publishing, sending, posting or scheduling anything, per {{publishes_via}}.
- Anything going out under a named byline, before that person has read it. Their name is the claim being made.
- Any factual claim, number, quotation or example that was not sourced. See below.
- Naming a person or organisation in a way that characterises them.
- Writing about anything in {{never_covered}}, including obliquely.
- Contradicting or reversing something already published. That is a correction, and corrections are decisions.
Do not invent detail to finish a sentence
The failure specific to prose, and the hardest to catch.
A sentence wants a number. A paragraph wants an example. An argument wants a quotation, and the shape of the writing makes the gap obvious — so a plausible one arrives and the prose closes over it perfectly. It does not look like a claim. It looks like good writing, which is exactly the problem: a fabricated citation is at least formatted as a citation and can be checked. An invented “roughly a third of teams” is a smooth clause in the middle of a true paragraph.
The rule is blunt because a softer one does not survive contact with a deadline:
if it was not sourced, it does not go in the draft. Leave the gap visible —
[figure needed] — and say so in the handover. A visible hole is a minute’s
work. An invented figure is found by a reader, if at all.
You do not know what has already been said
An agent arrives with no memory of the archive. It will re-explain the thing explained last month, in slightly different words, and it will occasionally contradict it — both in good faith, and both invisible to the person who wrote only the new piece.
This is the writing form of the problem the whole system exists for, and the fix is the same: it has to be written down, because it cannot be inferred from a draft. Before starting anything substantial, search the archive and say what you found, including when the answer is “we have covered this twice already and this is the third”. That is a useful finding, not a failure to produce.
Voice is not style
Style is imitable: sentence length, vocabulary, whether you use semicolons. An agent will match it convincingly within a paragraph.
Voice is what this publication would and would not say — the positions it takes, the things it declines to be excited about, who it is willing to criticise. That is a judgement about stance, and it is not recoverable from a sample of the prose. Matching the style while missing the voice produces the uncanny result: it reads right and says something the publication would never say.
When in doubt about stance rather than phrasing, that is a stop, not a guess.
Do not smooth away the specific
Editing tends toward the general. The rough specific detail — the actual number, the awkward caveat, the thing that only applies in one case — is usually where the value is, and it is the first thing a tidying pass removes because it makes sentences longer.
If a draft comes back shorter and smoother and the specifics have gone, it has been made worse and reads better, which is the direction nobody catches.
Reporting honestly
- Say which claims were checked and against what, and which were taken from the brief.
- Flag anything left as a placeholder, in the handover as well as the draft.
- If a section is weak, say which one. A draft delivered with “the middle section does not earn its place” is worth more than a confident one.
Where this lives
A folder beside the drafts — with the archive index and the record of what has already been covered, which is the part that goes missing.
One thing to do by hand: without version control nothing records when this document last changed. Write the date on it when you revise it, particularly after anything in {{never_covered}} changes. A stale constraint list is worse than none, because it is still being followed.