Editing AI Prose Into Something Publishable

Unedited machine prose has a small number of recognisable habits. Once you can name them you can remove them mechanically, in about twenty minutes for a thousand words. Here are eleven, each with a before, an after, and the rule that generalises.

The eleven tells at a glance

None of these is unique to machines. They are the habits of anybody writing to fill a space rather than to say something, which is why the same edits improve human copy. What is distinctive is the density: all eleven, in every piece, reliably. They are also concentrated in the rewriting stage rather than the diagnostic one, which is the argument for keeping the model on diagnosis rather than rewriting in the first place.

Four tells in the structure

1. The tricolon

Three parallel items is the most stable rhythm in English prose, which is why it appears whenever text is generated to sound finished. The test is simple: delete the third item and see whether anything was lost.

BEFORE
Good documentation is clear, concise and comprehensive.

AFTER
Good documentation is complete. Clarity follows from that; brevity does
not, and a short page that makes the reader guess has failed.

The rule. Wherever you find three, keep the ones that are doing work. If all three survive, they will usually need different amounts of space, which breaks the parallelism — and that is the point.

3. The restating summary

A final section beginning “In conclusion”, “To sum up” or “Ultimately” that tells the reader what they have just read. It exists because a generated piece needs a signal that it has ended.

BEFORE
In conclusion, we have explored three approaches to caching, each with its
own strengths and trade-offs. By understanding these differences, teams can
make more informed decisions about their infrastructure.

AFTER
Use the response cache unless your prompts share a long prefix, in which
case the prefix cache pays for itself at about four calls per prompt.

The rule. Delete it. If the piece genuinely needs an ending, the only thing worth putting there is what you have not said yet — the recommendation, the caveat, or the thing you would tell somebody in the corridor.

4. The uniform list

Six bullets, each a bolded two-word label, a colon, and one line of gloss. It looks organised. What it actually does is remove every signal about which item matters, because a reader infers importance from length and position.

BEFORE
- Scalability: The system can handle growing workloads.
- Reliability: The system remains available during failures.
- Security: The system protects data from unauthorised access.
- Maintainability: The system is easy to modify over time.

AFTER
- Security is the only one of these that cannot be retrofitted. An
  authorisation model chosen in week one decides what is possible in
  year three, and unpicking it means touching every endpoint.
- Reliability and scalability are the same problem here, because the
  failures we see are all load-shaped.
- Maintainability we are deliberately trading away until the schema
  stops moving.

The rule. Let the items be different lengths. Delete any that were there for symmetry. A list where one item is three sentences and another is four words is a list that has been thought about.

9. False balance

Two positions presented as equally weighted, followed by a sentence placing the truth between them. It is the safest possible thing to write and it transfers the entire decision back to the reader.

BEFORE
Some argue that fine-tuning is essential for domain performance. Others
maintain that retrieval is sufficient for most use cases. The truth likely
lies somewhere in between, and the right choice depends on your specific
requirements.

AFTER
Retrieval wins whenever the knowledge changes faster than you can retrain,
which for most internal documentation means weekly. Fine-tuning wins on
format and tone, where the target does not move. If your problem is that
the model does not know something, retrieve; if it is that the model will
not shut up, tune.

The rule. Replace “it depends” with what it depends on. If you cannot name the condition, you do not yet know the answer, and the honest fix is research rather than a hedge.

Four tells in the sentence

2. Not just X, but Y

The construction promises escalation and delivers a synonym. Variants: “this is not merely a change, it is a transformation”, “more than a feature, it is a philosophy”.

BEFORE
This is not just a refactor, it is a fundamental rethinking of how the
service manages state.

AFTER
The refactor moves session state out of the process and into Redis, so a
deploy no longer logs everybody out.

The rule. Delete everything before the comma and make the second half the sentence. If the second half then looks thin, that is the actual problem the construction was hiding.

6. The em-dash joint

The dash is a good piece of punctuation used once. Used as the default connector it removes the distinctions the other marks were carrying, and it is the single most commented-on surface feature of machine prose.

BEFORE
The API — which was rewritten last year — now supports streaming — a change
that many users had been asking for — and it is available on all plans.

AFTER
The API, rewritten last year, now supports streaming on all plans. Users
had been asking for it since the first beta.

The rule. One dash per paragraph at most, reserved for the interruption you actually want. Everything else becomes a comma, a full stop, a colon or a pair of brackets, and choosing between those four is most of what sentence rhythm is.

7. The participial tail

A sentence finishes, then a comma and an -ing clause arrive to explain what the sentence just achieved. “…, ensuring that…”, “…, allowing you to…”, “…, making it ideal for…”. It is the prose equivalent of a nod at the end of your own sentence.

BEFORE
The gateway retries on 429 with exponential backoff, ensuring that transient
rate limits do not surface to end users and improving overall reliability.

AFTER
The gateway retries on 429 with exponential backoff. A transient rate limit
therefore costs a user latency rather than an error.

The rule. Search for “, ensuring”, “, allowing”, “, making”, “, providing” and “, enabling”. Each one is either a real consequence, in which case promote it to its own sentence with a subject, or a restatement, in which case delete it. In practice about two-thirds are restatements.

8. Actorless abstraction

Verbs turned into nouns, with the person who did the thing removed. This is the register of a status report, and it is what makes competent machine prose feel like it was written by an institution.

BEFORE
Implementation of the new review policy resulted in a significant reduction
in time-to-merge across the organisation.

AFTER
We changed the review policy in March: one approver instead of two. Pull
requests now merge in about a day, down from three.

The rule. Find the verb hiding inside the noun (“implementation” → implemented, “reduction” → fell) and give it a subject who could in principle be wrong. Actors make sentences falsifiable, which is why the abstraction was there.

Three tells in the stance

5. Hedge stacking

One hedge is honesty. Three is the absence of a claim, dressed as caution. “May potentially”, “can sometimes help to”, “in many cases could be considered”.

BEFORE
Prompt caching may potentially help to reduce latency in some cases,
depending on a variety of factors.

AFTER
Prompt caching cuts time to first token when the cached prefix is long
enough to dominate prefill, and does nothing at all when the shared prefix
is shorter than the provider's minimum.

The rule. Keep at most one hedge per claim and convert the rest into the condition. “Sometimes” is not a hedge if you say when.

10. The announcement

A sentence whose content is that the next sentence is coming. “Let us take a closer look at…”, “It is important to understand that…”, “In this section we will explore…”. Headings already do this job.

BEFORE
Let us take a closer look at how caching works. Understanding the mechanics
here is crucial for anyone building on top of these APIs.

AFTER
A cache hit skips prefill for the shared prefix, so it changes time to first
token and leaves the generation rate alone.

The rule. Delete the announcement and start with the content. While you are there, treat “crucial”, “vital”, “robust”, “seamless” and “powerful” as deletion candidates: an adjective that asserts importance is doing the job a fact should be doing.

11. The unrequested call to action

The piece ends by telling the reader to begin their journey, start experimenting today, or choose the approach that is right for them. It is addressed to nobody, and it is the clearest signal that the text was produced to fill a slot.

BEFORE
Whatever approach you choose, the important thing is to start experimenting
and find what works best for your unique needs.

AFTER
(deleted)

The rule. There is no after. End on the last sentence that contained information. Readers do not need to be released.

The order to edit in

Order matters because early passes delete material that later passes would otherwise have polished. Doing it backwards is how a twenty-minute edit becomes an hour.

1. Cut whole sections. The summary (tell 3) and the call to action (tell 11) go first, along with any section that restates another. Expect to lose 10–20 per cent of the words here and none of the content.
2. Fix the structure. Break the false balance (9), un-uniform the lists (4), and thin the tricolons (1). This is where you discover whether the piece has an argument.
3. Add what is missing. Every claim that survived needs a condition, a number or an example. This is the only pass that adds words and it is the one that makes the piece worth reading.
4. Repair sentences. Tells 2, 6, 7 and 8, mechanically, with search. Fifteen minutes for a thousand words.
5. Fix the stance. Hedges (5) and announcements (10). Read the result aloud; the announcements are the sentences you skip.
6. Read for voice last. By this point the text is correct and anonymous. Voice comes back from specifics and from asymmetry, not from adding adjectives.

Several of these passes are mechanical enough to automate as a warning list rather than a rewrite — the style guide page ships the checker.

What editing cannot fix

This is the part the other guides leave out, and it decides whether the edit is worth doing at all.

• An absent argument. Every tell above is a surface feature. If the piece has nothing to say, removing the em-dashes produces well-punctuated nothing. The test is whether you can state the claim in one sentence that somebody could disagree with; if you cannot, stop editing and go back to the outline.
• Missing specificity. You cannot edit in a number that nobody has. Where the draft says “significantly faster”, the fix is measurement, not word choice.
• Fabricated support. A confident false citation survives every pass in this article, because it reads beautifully. Verification is a separate job and it is not part of a style edit — it belongs to the span-verified research workflow upstream and to the human gate downstream.
• Voice, if there was never any. Voice is accumulated decisions about what to leave out. It cannot be applied afterwards, which is why “make this sound more human” produces contractions and exclamation marks and nothing else.

A five-minute check before publishing

Six questions. If more than one answer is no, the piece is not ready regardless of how clean the prose is.

1. Can you state the claim in one sentence that a reasonable person could dispute?
2. Is there at least one specific — a number, a name, a version, a command — that the reader could not have guessed?
3. Has every fact, quotation and figure been checked against a source you opened yourself?
4. Does the piece end on information rather than on an exhortation?
5. Is there a paragraph that says what is not true, or where the advice stops working?
6. Search the document for “delve”, “tapestry”, “landscape”, “realm”, “testament to”, “navigate the complexities” and “in today’s”. Zero hits is the target, and each hit is usually attached to a sentence with no content.

The last one is worth a warning. Word-level tells shift as models change, and a checklist of vocabulary goes stale faster than a checklist of constructions. The eleven above are structural, which is why they have survived several generations of model; treat the word list as disposable and the construction list as durable.