2026-08-25 · Guide

Bots for Writers: Research, Outlines, and Repurposing

Only six hours of a writing week are actually writing

You publish two pieces a week. The part where sentences get made is maybe six hours of that week. The rest goes somewhere less flattering: opening forty tabs to find the four that matter, chasing a source who has not replied, checking whether you already made this argument in March, turning one essay into a newsletter and a thread and a post, and invoicing for the piece you filed nine days ago.

That distribution is the opportunity and the trap at the same time. The thirty hours are recoverable because most of them are retrieval, comparison, and format conversion, all of which have checkable outputs. The six hours are not recoverable, because they are the only part of the week a reader is actually paying for.

So the rule for a writer's roster is narrower than the rule for most roles. A bot is welcome anywhere its output is an input to you. It is unwelcome anywhere its output is the thing the reader sees. Every recommendation below sits on one side of that line, deliberately.

Staff six jobs, and keep four of them away from the reader

Six jobs worth handing over, with the catalog listing to adapt rather than start from a blank charter. Note that four of the six never produce a reader-facing word at all.

Two of these are enough to start, and the pair to start with is the source sweep and the beat watch, because they attack the biggest single block of non-writing time and their output is trivially checkable.

Delegate by who reads the output, not by how hard the task is

Most delegation advice sorts work by difficulty, which is the wrong axis here. A bot writing a headline is doing something easy and unforgivable. A bot reading four hours of transcript is doing something tedious and completely safe. The question that sorts correctly is simpler: when this output exists, who reads it?

Two rows get handled wrong in opposite directions. The headline gets delegated because it is short, and short reads like low stakes, when it is the line most readers judge you on. Source screening gets kept because it feels like craft, when it is mostly reading things to learn they were irrelevant.

The awkward case is the last row, because the answer changes halfway through. Repurposing starts as a job whose output you read and ends as a job whose output a reader sees, which is exactly where the boundary has to sit.

Research is the surface where a bot earns its keep

Research is retrieval plus triage. Both are verifiable in seconds, which is exactly the property that makes a job safe to delegate: you can tell a good output from a bad one at a glance, so you will keep checking in week six.

The format that makes it verifiable is a claim line that carries its own receipt. For every assertion, the bot returns the source name, the URL, the publication date, and the sentence it is standing on, quoted. Nothing else counts as research. A paragraph of fluent summary with no links attached is not a research artifact, it is a guess with good posture, and the failure is invisible precisely because fluent wrongness reads better than a hedge.

There is a practical wrinkle worth designing around. The runtime uses static egress IP addresses, and some services flag datacenter addresses, which means a research bot will occasionally be blocked from a page it can see the headline of. The right behaviour is a line in the report saying which sources it could not open. The wrong behaviour, and the default behaviour of any model asked for a summary, is to fill that gap from memory. Write the instruction explicitly: report what you could not read, never reconstruct it.

Structure is safe, sentences are not

An outline is a claim about the shape of an argument. You can read one in ninety seconds and know whether it is right, and even a wrong outline is useful, because disagreeing with a proposed structure is one of the fastest ways to discover what you actually think.

A draft paragraph does not have that property. A paragraph you half agree with costs more to fix than to write, because you have to unpick someone else's sentence rhythm before you can put your own back. This is the practical reason to keep bots at the outline layer, before any argument about voice.

Ask for the outline in a shape that exposes gaps rather than hiding them. For each section: the claim, the evidence in hand, the evidence still missing, and the strongest objection. The missing-evidence column is the part that changes your week, because it turns "I should research this" into three specific lookups you can finish before lunch.

Voice is the only part of this job that has not been commoditised

Every role has one risk that is specific to it. For a writer it is not accuracy, which is checkable, and it is not speed. It is that a bot drafting in your name produces prose that sounds like nobody.

Work through what a reader is actually buying and the stakes get clearer. Information is not it: the facts in your piece are in ten other pieces, usually free, usually faster. Structure is not it either, because a competent outline of any subject is now available to anyone who asks for one. What is left, and the only thing left, is the specific person doing the noticing. The judgement about which detail matters. The willingness to say the unhedged version. The rhythm that makes one paragraph land and the next one turn. That is what somebody subscribes to, and it is the one input that cannot be produced from a prompt because it is not a technique, it is an accumulation.

The economics point the same way. Competent, voiceless prose is now in effectively infinite supply, and the price of anything in infinite supply falls toward its marginal cost, which here is roughly nothing. Competing on that axis means competing where you have the least possible advantage. Meanwhile the supply of work that could only have come from one person has not increased at all, which is the only market position worth defending.

There is a second-order cost too. Voice is what makes you findable by the people who pay you: the editor who remembers a line, the reader who forwards a paragraph. Nobody forwards median prose. It is not bad, it simply never creates the occasion.

Voice goes gradually, which is why the guard has to be structural

The way it goes wrong is not dramatic, which is why people miss it. Generic prose is easy to approve. You do not reject the draft, you tweak it: swap a word, cut a clause, ship it. Do that twice a week for three months and the archive reads like a content mill, and there is no single piece you can point at where it happened. The gradualness is the mechanism.

Look closely at that moment of approval, because it is where the whole thing turns. Editing a competent draft is genuinely easier than writing a paragraph from nothing, and the result is genuinely publishable. There is no point at which you make a bad decision. You make the same reasonable decision eighty times, on eighty deadlines, each of them defensible on its own, and the sum of them is a body of work that could have been produced by anyone. The failure has no event in it, which means there is nothing for your judgement to catch.

That is precisely the argument for a structural rule rather than a careful habit. A habit is a decision you make when you are tired, and tired is when the draft looks good enough. A structural rule is a decision you already made when you were not tired, and it only has to hold once.

So the line is not "the bot must write well". It is: the bot may produce anything you will read, and nothing a reader will read, until you have written it yourself. Every listing in the botskills.sh catalog declares that kind of line as a boundary, and for writing bots the boundary is close to the whole design. If you want the phrasing that makes a boundary hold up under pressure, writing a boundary that actually constrains a bot covers it.

The rule has a useful side effect. Denied sentences, the bot has to hand you something else, and what it hands you is better: gaps, objections, the argument you already made in March and forgot. You end up with a research assistant rather than a co-writer, which is the role that was actually vacant.

Audit the drift with a test that can fail

Nobody catches gradual drift by intending to. Two checks catch it, and both are cheap enough to run quarterly.

The first is the byline test. Take five paragraphs from a piece you wrote before any of this, five from something you published last month, strip the titles and the dates, and hand them to somebody who reads you. Ask which set is which. If they shrug, you have your answer, and you got it from the only judge that matters, which is a reader rather than your own sense of the work.

The second is a checklist, because the tells are consistent. Prose that has drifted has a texture, and the texture has parts.

Run it against your own last three pieces before anything a bot touched. Half the time the drift predates the bot, which is worth knowing, because then removing the tool fixes nothing.

Repurposing without republishing yourself

Repurposing looks like the exception, because the source material is your own published work and the judgment involved seems mechanical. It is the safest generative task on the list, but it is not free, and it fails in a specific way: a thread version of an essay is not a compression of the essay, it is a different piece with a different argument shape, and a bot that treats it as compression produces something that is technically accurate and completely inert.

Three rules keep the queue honest. One derivative carries one idea, not a summary of five. Nothing goes out the same day as the original, because the derivative should be written against how the original actually landed. And every derivative links back, which keeps the incentive pointed at the work rather than at the platform.

The bot's job in all three is proposal. It finds the candidate, extracts the passage, and drafts a version you rewrite. Nothing about that requires publishing rights, so it should never hold them.

The research and outline charter, pasteable

This is the highest-value single bot for most writers. It runs overnight against the topic you queued, and you read it with coffee.

ROLE
You are my research and outline assistant for a working writer.
You produce material I read. You never produce material a reader reads.

TRIGGER
Every weekday at 06:00, process the top item in my "queued topics" note.
If the queue is empty, do nothing and say so in one line.

INPUTS
- The queued topic and my one-line angle for it
- Public web sources published in the last 24 months, plus primary
  documents of any age (filings, specs, transcripts, papers)
- My own published archive, to flag where I have covered this before

OUTPUT, in this order
1. PRIOR ART: any piece of mine within one topic of this, with the date
   and the argument I made then. If I have contradicted myself, say so.
2. CLAIM TABLE: 8 to 15 rows. Each row is one factual claim, the source
   name, the URL, the publication date, and the exact sentence it rests
   on, quoted.
3. COULD NOT READ: every source I asked for or you found that would not
   open, blocked, paywalled, or timed out. Name it. Do not fill the gap.
4. OUTLINE: 5 to 7 sections. For each: the claim, the evidence from the
   claim table by row number, the evidence still missing, and the
   strongest objection a knowledgeable reader would raise.
5. OPEN QUESTIONS: the three things I would need to ask a human to know.

RULES
- No claim without a row in the claim table. No row without a URL.
- Never paraphrase a source you could not open.
- Never write body prose, an intro, a conclusion, or a headline.
- Quote at most one sentence per source, in quotation marks, attributed.

BOUNDARY
You never publish, post, send, or schedule anything, anywhere, ever.
You never contact a source, an editor, or a reader on my behalf.

The third output block is the one people delete because it looks like an admission of failure. Keep it. A named gap costs you one lookup, and a silently filled gap costs you a correction with your name on it.

Expect the first month to be mostly charter edits

The first week is not the bot working, it is you finding out what you actually meant. Here is the shape of it.

Day one, the 06:00 output is too long and slightly wrong. The claim table has fourteen rows, four of which are the same fact from four aggregators repeating one press release, which is the characteristic failure of a first run. The outline is competent and generic, because the angle line you queued was "write about X" rather than an argument. You spend twenty minutes reading it and get one usable thing: a paper from 2023 you had not seen.

Week one is charter work. You add a rule that primary sources outrank coverage of them, and that four rows citing the same original count as one row. You start writing the angle line as a sentence with a verb in it, which changes the outline more than any other single edit you will make. You notice the "could not read" block is empty every day and get suspicious, so you queue a topic you know has a paywalled source, and check.

By week three the pattern is stable and the value has moved. You stop reading the outline first and start reading the prior art block, because being told what you argued fourteen months ago is the output nothing else produces. The claim table becomes a lookup rather than a read: you scan the URLs, open two, and get on with it.

By week four you will know which of two states you are in. Either the outline is something you write from, in which case the setup has paid for itself several times over, or it is something you skim and abandon, in which case the fault is almost certainly the angle line and not the charter. Fix the input twice. If it still fails, retire the bot and keep the source sweep, which almost never disappoints because retrieval is the part machines are genuinely better at.

Recognise the five ways a research bot fails

None of these look like failure while they are happening, which is the point of listing them.

There is a sixth that has no row because it has no symptom you would notice: you stop reading the output. Nothing breaks, the file still arrives at 06:00, and the cost carries on. Retire it the week you notice, not the month after.

What not to automate if you write for a living

This is the opinionated part, and the list is shorter and harsher than the roster above.

The sentences. Not the first draft, not the "just get something on the page" draft, not the newsletter because it is only the newsletter. The blank page is the job.

The thesis. A bot proposing angles is a research tool. A bot deciding which angle is right has taken the one decision that makes the piece yours.

Contacting a source. Never let a bot email an expert, a PR contact, or an interviewee. A source's willingness to talk to you is a relationship, and it is also the thing you burn first if a machine sends them a slightly wrong message signed with your name.

Pitches. An editor reads your pitch as a sample of your writing, because it is one. This is the single most expensive place to sound like everyone else.

Reading the primary thing. The book, the transcript, the filing, the report. A summary of a summary is where wrongness enters and it never announces itself.

Corrections. When you get something wrong, the correction is a piece of character. It is not a formatting task.

If you are running a whole business alone and want the surrounding setups rather than just the writing ones, the one-person company guide has the non-writing half.

Answer the volume argument, because it is the real one

The strongest case against everything above is not that machines write well. It is that they write enough. Somebody on your subject publishes five times a week to your two, and if attention follows volume, keeping the sentences is a principle you pay for in reach. That deserves better than dismissal.

It wins in specific places, and they are identifiable. Where the value of a piece is that it exists rather than that it is good: reference material, release notes, roundups, coverage where being first matters more than being read. Where the reader wants an answer and does not care who is speaking. Where consistency across hundreds of pages is the actual product, which is what good documentation is. In those settings a distinctive voice is not an asset, it is noise, and the machine is the correct tool.

It loses where the reader chose you specifically. Essays, criticism, reporting, newsletters people opted into, anything somebody reads because of the byline rather than despite it. In that market, doubling output with prose that sounds like everyone does not double reach, it dilutes the reason anyone subscribed. The competition is not the other writer publishing five a week. It is the enormous and growing supply of adequate pieces on your subject, and adding to that supply is not a way to stand out from it.

There is also a timing asymmetry worth being clear-eyed about. The volume play gets weaker as more people run it, because it depends on being unusual. The voice play gets stronger as more people run the volume play, for exactly the same reason. If you are choosing a strategy for the next three years rather than the next quarter, those two curves are pointing in opposite directions.

And one honest exemption. If you ghostwrite, the voice you are protecting is someone else's and the calculus changes: your job is fidelity to a person who is not you, which is a different skill and a different set of checks. The byline test still applies, you just run it against their archive rather than yours.

Measure two things, and retire the bot if they fail

Two numbers, both cheap to keep, both worth more than any impression you have about whether the bot is helping.

First, the link check. Take ten claim rows at random from a week of output and open every URL. You want ten for ten on links that exist and say what the row says they say. At nine you have a tuning problem. At eight or below the bot is a liability, because you are now the fact checker for a machine that is supposed to be saving you time.

Second, the outline-use rate. Across ten pieces, how many outlines did you work from mostly unchanged, versus discard? Below half is fine at the start and not fine by piece thirty, and the fix is almost always the angle line you feed in rather than the charter.

Watch the cost side too. A subscription includes a weekly usage allowance and overflow is billed on demand from model and token cost, so a research bot producing output you stopped reading in week three is worse than no bot: it costs money and it makes you feel covered. If nobody reads it, retire it.

Keep reading: What Is a Grok Bot? The Plain Explanation for Non-Engineers, Bots for Consultants, Bots for Sales Reps.

Frequently Asked Questions

What can AI bots actually do for writers?

The reliable wins are retrieval and triage, not composition. A bot can sweep long interviews and talks into timestamped notes, watch the pages and publications on your beat for changes, propose angles with the crowded ones flagged, track what is drafted versus filed versus overdue, and propose derivatives of work you already published. Each of those produces something you read and act on, and each is checkable in under a minute. The composition itself, meaning the actual sentences a reader sees, is where the value of a writer now sits and where delegation costs the most.

Should a bot write the first draft?

No, and the reason is practical before it is romantic. A paragraph you half agree with is more expensive to repair than an empty page, because you have to unpick someone else's sentence rhythm before your own will fit. The blank page is where you find out what you think. Have the bot deliver an outline with claims, evidence, gaps, and the strongest objection to each section, then write from that. You keep the speed benefit of not starting cold and you keep the part of the job that is actually yours.

How do I keep an AI bot from flattening my voice?

Draw the line structurally rather than relying on judgment in the moment. The bot may produce anything you will read and nothing a reader will read, and that rule goes in the charter as an explicit boundary on publishing, posting, and sending. Voice loss is gradual and comfortable: generic prose is easy to approve, so you tweak rather than reject, and three months later the archive has drifted with no single piece to blame. A structural line survives a tired Thursday. An intention to edit carefully does not.

Can a research bot be trusted with citations?

Only if the format forces the receipt. Require every claim to arrive as a row carrying the source name, the URL, the publication date, and the exact quoted sentence it rests on, and require a separate block listing every source that would not open. The runtime uses static egress addresses and some sites refuse them, so blocked pages are normal and a bot asked for a summary will otherwise fill the gap from memory. Then verify: pull ten rows a week at random and open every link. Ten out of ten is the bar.