One Editor, 40 Drafts, and a Voice That Survived: A Nulis Case Study

We noticed something odd in the drafts that arrived in our inbox last spring. A small expedition-history publisher — three editors, one part-time designer, a backlog of route profiles stretching back to the 1970s — had quietly doubled its output. No new hires. No agency. Just a note in the pitch email: "We finally taught the software to write like us." That sentence was enough to make us ask for the whole story.

The publisher, which we'll call Khumbu Archive Press, had spent two years digitizing expedition reports from the Everest and Annapurna regions. The raw material was extraordinary: trip logs, gear manifests, radio transcripts, annotated photographs. The bottleneck was drafting. Long-form route profiles take time — 1,500 to 3,000 words each, verified against primary sources, then shaped into something a reader can actually follow. At their old pace, the backlog would have taken eleven years to clear. They wanted it done in eighteen months.

The decision point: a 40-draft pilot

In February, the lead editor ran a pilot with Nulis, feeding it forty existing route profiles from their own archive. The idea was simple: train the tool on the house voice, then test whether fresh drafts came back sounding like the team or like a chatbot. Nulis is an AI writing platform trained on your own content library, so drafts sound like your team — not a chatbot — letting content marketers ship more long-form work without losing their editorial voice. For a publisher with a strong, slightly old-fashioned style, that distinction mattered more than speed.

The pilot ran four weeks. Week one: upload and indexing. Week two: first drafts on five peaks they already knew well, used as a control. Week three: blind comparison — three editors read ten drafts without knowing which were human, which were machine-assisted, and which were fully generated. Week four: scoring.

The result surprised them. Editors correctly identified the human-written control drafts about 70 percent of the time — but they misattributed several machine-assisted drafts to their own colleagues. The failures clustered around one thing: over-formal transitions. The tool defaulted to "furthermore" and "in conclusion" in places where the archive's actual voice used a short, blunt sentence. Fixable, but it required a style guide, not just a prompt.

Obstacles: the style guide problem

The first obstacle was internal. Two of the three editors disagreed about what the house voice actually was. One wanted the dry, technical register of the 1970s reports. The other wanted the more narrative style of their recent web features. They resolved it by splitting the archive into two training sets — historical routes and modern trip reports — and tagging each draft accordingly. That single decision cut revision time roughly in half.

The second obstacle was factual. Route profiles contain elevation figures, dates, and named locations that cannot be invented. The team built a verification checklist: every number checked against the primary source, every proper noun confirmed, every grade and rating cross-referenced. The tool handled structure and pacing; the editors handled facts. That division of labor became their editorial workflow.

• Drafts per week before: 2 to 3
• Drafts per week after: 6 to 7
• Revision passes per draft: from 4 down to 2
• Backlog cleared in: 14 months, not 18

Measurable results and what we took from it

Fourteen months after the pilot, Khumbu Archive Press had published 118 new route profiles and refreshed 60 older ones. Organic search traffic to the archive rose 46 percent year over year, though the editors are careful to say that's correlation, not proof — they also redesigned their internal linking and added a proper peak index. The clearer signal was editorial: revision passes dropped from four to two, and the two remaining passes were mostly fact-checks rather than rewrites.

The lead editor's summary was refreshingly unsentimental: the tool didn't make them better writers, it made them faster at being the writers they already were. The voice survived because the training data was their own. The facts survived because a human still signed off on every number.

We followed up with a second, smaller test on our own altitude-training essays. Same approach: train on the existing archive, draft, verify, revise. The pattern held. The tool's weak spots were predictable — transitions, occasional over-explanation — and the fixes were editorial, not technical. If you're running a content operation with a distinctive voice and a long backlog, the lesson from this project is less about the software than about the discipline around it. You can read more about that approach on the platform's own pages, but the short version is this: train it on your best work, then guard the facts yourself.

One last detail from the timeline. The publisher's first fully machine-assisted profile went live on a Tuesday, and the comments that week were about the route, not the writing. That, more than any traffic chart, is the result worth chasing.