05 · Editorial agents

Leadership wanted better drafts. The editors needed evidence.

The first interviews changed the product in the best way. Leadership asked for faster drafts; editors showed us they needed evidence for a better conversation before drafting began. So the first product checks the outline and leaves the editorial call with them.

45 of 165facts were present but did not support the argument

My roleProduct lead; built the plugin's skills, co-developed with an engineer
WhenJun 2026 – present
Userseditors, writers, and research analysts
StatusIn pilot on live reports

A research institute wants to be one of the first agent-enabled think tanks in the world. Leadership's ask was clear: better first drafts, faster. Then I interviewed the editors, and two things came out. They were afraid of being replaced. And the pain started long before drafting: storylines arrived at wildly different maturity, and editors lacked the evidence to push back on author teams, who are mostly analysts pulling facts together, not writers. Drafting was the easy part.

I'm the product lead for the institute's editorial AI work, and I'm hands-on-keyboard: I built the skills in the plugin myself and co-developed the plugin with an engineer. I interviewed eight editors and writers, set the order we built in, and wrote the team's charter.

Keep the human where their talent is. Editing is subjective, and an editor's real skill is shaping an argument and influencing the people who wrote it. So we were careful about where AI helps and where the editor stays in charge: the tools find evidence, and the editor makes every call.

Fix the outline before the draft. Leadership asked for drafting help. The interviews said the pain started upstream, so the first build was an outline check: it reads a storyline against its evidence, lists the facts that support none of the points, and flags inconsistencies and repetition, keeping the editor's words exactly as written. The drafting skills came second. The first logged ruling: flag, never silently correct.

  1. storyline + evidencethe author team's outline and its facts
  2. outline checkorphan facts, inconsistencies, repetition
  3. independent checkrules or a second model; nothing checks its own work
  4. the editor rulesevery ruling logged
  5. draftsstoryline to draft, charticle, op-ed, regional cut, executive summary

Every output is checked by something that didn't write it, and nothing moves until a person says so.

Storyline checkmade-up report: "Why neighborhood libraries are busier than ever"

Governing thought, ¶2: Libraries are busier because they've become the city's free third place.

  1. Visits are up even as loans fall6 supporting facts
  2. People come for the space, not the books4 supporting facts
  3. Small branches see the biggest jump5 supporting facts

supports no pointCard fees were dropped in 2019.

supports no pointThe average branch is 41 years old.

inconsistent¶4 says visits rose 18%; ¶9 says 12%.

repeatsPoint 2 shows up again in ¶11.

Flags only: the check never rewrites a sentence or approves anything.the editor decides
A made-up report run through the same check: the main points, the facts behind each one, and the facts that support none of them. The editor rules on every flag.
Built withOutline checkReader testDrafting skillsIndependent verificationDecision log
45 of 165
facts in one live outline that supported none of its points, found in minutesthe check's own output on a real storyline
< 1 wk
per proof of concept, each built on live material
~99
agents people had already built across the institute; we reused their best parts in the plugin, so the work downstream gets easier

Tested on live reports, not demos. Every proof of concept ran on real material, and the editors tried the check on their own work.

I love it when it spots these.An editor, testing the outline check on a live report
This would be useful for me right now.The same editor, four days later, on the orphan-fact list

The editors keep the judgment. The tools make it visible, and the log makes it defensible. Several tools we inherited stopped being necessary once the outline got better, so we retired them and moved the fix upstream. The plugin is in pilot on live reports.

  • Meet people where they are. A skill for the moment earned the right to build the plugin for the long run.
  • Time saved isn't the metric. Editors told me what matters is feeling effective: walking into the author meeting with evidence.