Want to know if this job is worth applying to?
San Francisco, California, United States
Onsite
Harper is an AI-native commercial insurance company in San Francisco. We're not bolting AI onto insurance — we're rebuilding the entire business as software, on a simple bet: turning expert human judgment into compute is one of the largest transitions left to make, and a trillion-dollar industry still run 90% by hand is the place to prove it. We've grown ~100x in the last year and we move at that speed — on-site, in person, long days, very high standards. Almost no one joins Harper for insurance; they join to build the company that replaces how it works.
You turn the judgment locked inside Harper's best operators into AI-legible knowledge — living docs, decision logs, and retrievable skills the agents can actually call — and you get the rest of the company running against it.
AI doesn't understand a company by default. It works only when the business is documented clearly enough for a system to retrieve the right context, recognize the workflow, handle the edge case, and escalate when human judgment is required.
Right now most of how Harper operates lives in people's heads: how a top rep sequences quotes, how service handles the weird bind, how market routing actually works, what a customer means when they push back. That holds at small scale. It breaks at ~1,000 new customers a month. Every undocumented process is a future failure mode; every AI-generated playbook that dies in a chat thread is throughput left on the floor.
The next bottleneck here isn't engineering. It's knowledge — and how fast people can absorb it. This role removes that bottleneck.
Be clear about what this is not. This is not corporate L&D. No LMS, no slide decks, no e-learning project, no making-the-Notion-pretty. This is knowledge engineering: sit with operators, extract how they actually think, and turn it into structured knowledge a human and a model can use.
Two tracks, running in parallel, at the intersection of Operations, Engineering, and RevOps.
Build the operating memory.
Embed with sales, intake, service, placements, and renewals. Sit with operators, listen to calls, shadow workflows, and document what people "just know."
Turn transcripts, Slack threads, Looms, and one-off explanations into source-of-truth docs, decision logs, playbooks, process maps, glossaries, and system-boundary docs (what each internal system does, where one stops and the next begins, and what gets misread).
Graduate stabilized rules into skills the agents can call, and partner with engineering on refresh automations so docs stay alive instead of going stale.
Hunt the edge cases — reworks, escalations, stale quotes, market follow-ups, binder/payment gaps, customer confusion — and write them down before they bite again.
Get the company to run against it.
When a rich AI-generated playbook lands, distill it into an executable plan: named owners, the first three moves, a rollout cadence and date. The plan doesn't run itself; you make it run.
Build the onboarding paths and setup scripts that get a new hire into Cursor, Claude Code, and the harness within a week.
Run cohort rollouts, drive adoption, and make activity visible — we should know who's actually in the harness.
Shape meetings in real time so they produce useful artifacts: decisions, owners, definitions, edge cases, open questions, next steps.
When the same problem shows up three times, turn it into a playbook, a QA check, a skill, or a product requirement. You'll work directly with the CEO when extraction calls for it.
An exceptional writer and synthesizer. You can take a messy transcript to a clear operating doc, and a clear doc to something a team actually executes against.
AI-native in practice — this is the bar. Not "I use ChatGPT." You have taste for when an output is structurally wrong, not just stylistically off. You prompt for extraction (decisions, contradictions, owners, edge cases), not just summarization. You know good output depends on good context, and you engineer that context upstream. Cursor, Claude Code, MCP servers, agent memory files, Granola, structured prompting are tools you reach for without thinking.
Genuinely curious about how organizations work, and happy to sit with operators to find out.
Structured but not bureaucratic. You care whether documentation changes behavior, not whether it looks polished.
Able to operate in chaos without becoming chaotic. Low-ego, persistent, and allergic to "someone should probably document that."
Backgrounds that can work (the skill matters more than the title): modern enablement at an AI-native company, qualitative/academic research, ethnography, instructional or curriculum design, knowledge management, product ops, technical writing, research ops, implementation, chief of staff roles, library and information science, AI ops / human-in-the-loop work, or startup operations where the systems were messy and someone had to make them legible.
Requirements
3–8 years in a relevant field (see above).
Exceptional written communication and strong information-architecture instincts.
Real AI-tool fluency in practice: Cursor, Claude Code, MCP servers, agent memory files, Granola, structured prompting, AI-assisted synthesis.
Demonstrated ability to interview stakeholders, extract operational detail, and turn messy conversations into clear decisions and source-of-truth docs.
A track record of running an adoption rollout that actually changed how a team worked.
Comfort in a fast-moving, ambiguous startup. Based in San Francisco or willing to relocate.
Nice to have: experience at an AI-native or dev-tools company; authoring Claude/agent skills or similar capability modules; RAG/search systems, data labeling, evals on knowledge systems, or human-in-the-loop QA; working with engineering on living-doc refresh automations; translating operator feedback into product requirements; taxonomy/metadata/content governance; insurance, fintech, B2B services, or another high-volume operational environment.
On-site in San Francisco, in the building with the operators whose knowledge you're extracting — this work doesn't happen remotely. The hours are long and the standards are high, because a rebuild this large doesn't happen part-time. The people who thrive here want it that way. If you want to walk into ambiguity, find the hidden logic, and turn it into systems the whole company runs against at AI speed, this is the seat.
Salary: $110,000–$170,000 + performance bonuses & equity
Location: San Francisco, in-office hours matching the rest of the company.
Schedule: Monday–Friday, in-office hours that match the rest of the company. The hours are long.
Benefits: Uber commuter benefits; breakfast, lunch, and dinner provided; snacks/drinks/coffee daily; free gym membership; health, dental, and vision insurance.
One to two screening calls — alignment on mission, pace, and role fit.
On-site super day — sit with operators, review transcripts, meet product and engineering, and show how you think.
