The Content Strategist: Directing the Agents

Ryan’s campaign plan reaches Naomi Feldstein on the Wednesday of week five. It is not a list of things to write; it is a matrix of journey stage against channel, with volumes and dates in each cell — fifteen pieces doing the unaware job by week six, a five-email nurture sequence, landing pages for four channels, the webinar programme, the gated research, the booth material. Counted properly, with channel versions and test variants, it comes to about three hundred assets, all of it drafted by the end of week eight.

Two years ago that requirement would have been refused, or met with an agency and a budget line. Naomi has content agents, and the honest position is that they could produce a rough first draft of all three hundred by Friday. She spends the first two days writing briefs instead — to anyone watching, the least productive week of the campaign, and the reason the other thirteen work.

What lands on Naomi’s desk

Ryan’s plan tells her what is needed, in which week, for a reader in which state of mind — and that last part is the useful part, because fifteen pieces doing the unaware job is a different order from fifteen product pages.

Camille’s messaging brief from module 5 tells her what may be said: one sentence, you hear it from us, not from your customer, three supporting points with their proof and sample sizes welded on, and two lists — words to use, words to avoid, each with its reason. Behind it sit Wes’s product truth document from module 2, Anika’s 340 verified market claims from module 3 with a URL and a date on each, and Joel’s interview verbatims from module 4, the only material in the building written in the buyer’s own voice.

And one date she cannot move. Everything she produces must clear Miriam Adeyemi’s review in week ten, for which Ryan has budgeted one week and written down, in the plan, that he believes the estimate is soft.

What a content strategist actually does

A content strategist commissions and edits: specify a piece, find someone to write it, decide whether the result does its job. The specifying is where quality is decided, because a writer who has understood the reader and the promise produces a usable draft, and a writer who has not produces something that reads well and points nowhere.

Agents change who the commissioning is aimed at and nothing else. Naomi still specifies and still checks — but she now specifies to a system that never asks a clarifying question and never comes back on Thursday to say the premise is wrong. Everything a good writer would have pushed back on must be anticipated in the brief or caught in the edit.

The software on Naomi’s desk

The brief template is the highest-leverage artefact on this desk, because agent output quality tracks brief quality almost exactly. It carries the reader, the journey stage, Camille’s sentence, Wes’s approved facts with sample sizes, the words to avoid and the job the piece has to do — and the same template makes human writers better, which is why teams that already wrote good briefs adopted agents easily and teams that did not, did not.

The drafting agent does what it is good at, at a volume that is genuinely new. The review queue is where the consequence shows up, and Naomi’s contribution is structural rather than editorial: claims are checked against the approved list before a draft leaves her desk, not after it reaches legal. The CMS publishes only from the approved library — which sounds bureaucratic until you have watched a withdrawn asset reappear on a staging page.

The four decisions

What a good brief to an agent looks like

The quality of what comes back from a content agent is set almost entirely before the agent runs. Nothing else Naomi does in these four weeks has the same leverage, and the gap between a good brief and a poor one is not ten per cent — it is the difference between a draft she edits and a draft she throws away.

Start with the poor one, because it is what most teams type: write a landing page for Supply Signal. What comes back is fluent, confident and useless. It addresses a reader already looking for supply chain software — the reader Camille spent module 5 explaining does not exist in this segment. It reaches for the category sentence, complete visibility and end-to-end risk management, because that is what the internet says about this category and the model has read the internet. It will very likely assert something nobody at Cadence approved. And it is 900 words where the placement needs 450.

Naomi’s brief for the same page runs about 350 words to produce a 450-word asset, and that ratio is not waste. It names the reader: a plant operations director at a manufacturer between $50m and $500m of revenue who has never heard of Cadence and is not looking for a product. It names the state: the page follows a LinkedIn ad about the cost of the status quo, so the reader arrives believing the problem might be manageable and believing nothing else. It carries Camille’s sentence verbatim and the approved facts in their approved wording — 1.4m suppliers, nine risk signals, live in six weeks, no native SAP connector until the following quarter — each with its sample size, inside the brief rather than in a footnote somebody may not carry through. It carries the words to avoid with their reasons, because a prohibition without a reason gets worked around by anyone under deadline. It fixes format and length. And last, the part most briefs never contain, it states the job: make the reader want the gated research enough to give a work email address — not book a demo, which is the wrong request at this stage and would fail silently.

The test of a brief is whether two competent writers, given it and nothing else, would produce pieces that do the same job — a test with nothing to do with agents and everything to do with what a good agency has always demanded of a client. Which is why teams that already briefed well adopted agents in a fortnight, and teams that did not got faster at producing material nobody had specified.

Volume, and the arithmetic of review

The agents deliver. A first draft of the full package that would have taken six weeks of writer time takes two. That is not a claim about tooling; it is what happens.

What the volume buys is real, and it is not “cheaper content”. It buys testable variants — five subject lines rather than two, enough to learn something rather than enough to say a test was run. It buys channel-specific versions instead of one message resized. And it buys iteration: when week-one data shows the unaware pieces working and the interested pieces not, the rewrite is days rather than a change request nobody has budget for.

Now the cost, because almost nobody does the sum. In the old world a launch of this size produced perhaps 90 assets. Two writers and an agency spent six weeks on them — call it 240 hours — and Naomi’s editorial pass ran about 45 minutes an asset, around 68 hours. Review was roughly 22% of the effort, which is why nobody ever managed it as a constraint. Now the matrix asks for 300 assets, drafted in two weeks. Briefing, directing and generating them costs perhaps 90 hours. But review has not changed at all: it is still 45 minutes to read a piece properly, check what it asserts and decide whether it does its job. Three hundred assets is 225 hours.

The totals barely move — about 308 hours before, about 315 after. What has moved is what the work is. Reading has gone from a fifth of the job to roughly seven-tenths of it, compressed into a shorter calendar. The gain is elapsed weeks, not hours; the hours were relocated from writing to checking, and checking cannot be handed back to the machine that created the volume.

So Naomi does what her own arithmetic tells her, which is the least glamorous decision in the module: she takes about sixty variants out of the matrix before a word is drafted. Not because they would be bad, but because nothing would be learned from them. They exist because generating them is free. Reading them is not.

She is also the first person in the chain to feel the pressure Miriam inherits in module 9, and one difference between their positions is worth naming. Naomi can decline to make an asset. Miriam cannot decline to review one.

The claim that keeps coming back

By the second week of drafting, one sentence has appeared in variant after variant: predicts disruptions three weeks early, sometimes in those words, more often in words that mean it.

It matters that this is mechanical rather than mischievous. The agent has no intent and no memory of the meeting in module 2 where Wes refused the verb. It has the source material — and in that material the Calder Thermal pilot statistic, seven of nine disruptions flagged at least three weeks early, is the only performance evidence in existence. It is the strongest sentence available. A model asked for persuasive copy produces the most plausible continuation of persuasive copy, and the most plausible continuation of a headline is a confident one. It finds the strongest supported-looking claim in the pack the way water finds the low point in a floor.

Camille’s words-to-avoid list catches some of it, and it is necessary. It is not sufficient, for two reasons. A prohibition operates on words while a claim is an idea: ban predicts and back come know before it happens, see it coming, get ahead of it. The list catches the token; the meaning walks round it wearing a different coat. And thirty headline variants is thirty independent attempts — a filter that catches nineteen in twenty still lets one through, and one is all it takes to reach print.

So Naomi’s fix is structural rather than a stricter list. The approved-claims list goes into the brief as positive material: the exact sentence Camille approved, n = 9, one customer, one quarter attached, presented as the thing to say rather than one more thing not to say. A prohibition leaves the strongest available claim where it was; an approved list changes what the strongest available claim is. Constrain by supply, not only by rule.

Then every draft gets a claims pass before it leaves her desk — a separate read whose only job is to check each assertion against the list, held apart from the editorial read, because a person doing both at once does neither. It belongs here rather than at Miriam’s desk for a plain economic reason: caught now, a claim costs a rewrite; caught in week ten it costs a rewrite plus a review cycle plus a slot in a week with no slack. And it does not catch everything — two ad variants reach Miriam in module 9 with the verb drifted back to predicts. Out of hundreds of assets that is a good hit rate and not a clean one, which is exactly why the second check downstream still exists.

What she will not delegate

Some pieces Naomi does not send to an agent at all, and the reasoning is a judgement rather than a rule.

The Calder Thermal customer story is the clearest case. Its whole value is that a named person at a real company said something in their own words about something that actually happened to them. An agent could draft it — quickly, plausibly, possibly with better sentences — and the words would not be theirs. What the reader is buying is not prose quality; it is testimony. So Naomi does the interview herself, writes it herself, and sends it to Calder to change whatever they want changed, which they will have to sign in writing anyway before module 9 lets it out of the building.

The founder’s launch post is the same argument in a different suit. It carries a byline, and a byline is a claim about who wrote it. A reader who later discovers a personally-voiced post was generated does not conclude that Cadence is efficient; they conclude the warmth was manufactured, and apply that backwards to everything else they have read. The shape of the judgement is this: where authorship is part of what is being claimed, authorship has to be real — anything in the first person, anything expressing a point of view, anything whose persuasive force depends on a reader believing a person sat down and meant it.

And Naomi is candid that this line moves. Two years ago she would not have delegated a subject line, an event follow-up or a webinar abstract; she now does all three without thinking, and the results are better. She expects to move it again — which is why she writes down where it sits today and the reason for each item on it, so that when it moves it moves as a decision somebody made rather than by attrition at four o’clock on a deadline.

The last twenty per cent

Not because the material is bad, but because plausible-but-wrong is a far more expensive defect than obviously-bad. A weak human draft advertises its own faults: the clumsy paragraph is where the writer did not understand the product. An agent draft has no such texture — the sentence that is subtly untrue reads exactly like the sentence that is fine. Nothing is signposted, so every line has to be read at full attention, which is slower per page than editing a human draft even when fewer changes are needed.

Then there are the fusions, where a draft takes two true facts — 1.4m suppliers watched, nine risk signals monitored — and produces a third nobody wrote and nobody can support: the most comprehensive supplier coverage in the category. Repairing that is not a copy edit; it means going back to the truth document and rewriting the argument the paragraph was making. Three sentences of work behind one adjective.

And there is the trap in the number itself. A draft that is 80% right feels nearly finished, so teams allocate a fifth of the remaining time and spend well over half. Naomi’s rule is a reframe rather than an estimate: an agent draft is not a draft approaching completion, it is a finished-looking proposal that nobody has checked.

Where this goes wrong

The commonest is making more of what was already not working. A team points agents at its existing content calendar and triples the output of pieces nobody read the first time. Every production metric improves, nothing downstream moves, and because output is what gets reported upwards it takes two quarters before anyone connects the two.

The second is quieter and worse. A reviewer handed three times the volume in a third of the time does not refuse; they begin, without announcing it and often without noticing, to sample. Nobody says “we have decided to check less”, and the first evidence is an asset in market that no human would have approved had a human read it.

What Naomi hands on

About 240 assets go to Theo Alvarez in module 8 and Miriam Adeyemi in module 9, and what makes the package unusual is the sheet attached to it: every factual assertion tagged with where it came from and what its status is. Approved traces to Wes’s truth document or one of Camille’s proof points, in the approved wording, sample size intact. Sourced traces to a line in Anika’s dataset with its URL and date. Net new appeared during drafting and traces to nothing — Naomi flags eleven of these rather than deleting them, because two or three are good ideas that need evidence somebody has not yet gone and found.

Theo gets the copy and the same sheet, because Camille’s point in module 5 applies to pictures as much as sentences: an image is not exempt from a claim merely because it contains no words. Miriam gets an early batch in week seven, as Ryan asked — the nurture emails and the gated research, the material that is stable early. The bulk still lands in week ten, because ads and landing pages cannot be finalised until Theo’s creative is paired with them in week nine.

And the limit of the tagging, which Naomi puts in the covering note rather than leaving to be discovered: you cannot tag what you did not recognise as a claim. The eleven flagged items looked like assertions. The dangerous ones looked like writing. Closing that hole is Miriam’s decision in module 9, and it takes a harder rule than a covering note.

Designing this desk’s agent: the drafting agents

Naomi’s agents are the ones people picture when they imagine AI in marketing. They are genuinely good, and their design problem is not quality but accountability at volume: three hundred assets, each containing sentences that assert things.

One concrete design choice. Require a claim manifest on every asset, with a net_new_claims array that must be empty before the asset can move to review. Anything the agent asserted that is not on the approved list surfaces automatically instead of being discovered by a lawyer three weeks later.

{
  "asset_id": "eml-nurture-03",
  "brief_id": "brf-mid-market-02",
  "claim_ids": ["clm-nine-signals", "clm-impl-six-weeks"],
  "net_new_claims": ["catches disruptions three weeks early"],
  "status": "blocked_pending_claim_review",
  "drafted_by": "agent",
  "directed_by": "naomi.feldstein"
}

The metric to track. Net-new claims per asset, which should trend toward zero as briefs improve, and rework rate at review. Both measure the same thing from different ends: how much of the speed gained in drafting is being handed back in checking.

Brief, or prompt?

Read each one and decide what Naomi would do, then tap a card to check.