Research labs are among the most documentation-intensive organizations I work with. For every hour of bench work, there's a corresponding obligation: protocol write-ups, progress reports, IRB amendments, poster abstracts, lay summaries for non-specialist audiences, annual reports to funders, literature review sections, and — most expensively in terms of senior researcher time — grant applications.
Grant writing is a skilled activity. It requires understanding the agency's priorities, knowing how to frame significance and innovation, writing to the scoring criteria. None of that is trivial. But a substantial portion of any grant document is not skilled writing. It's structured recounting of work that's already been done, in a format that's largely standardized, applied to a research context the PI knows cold.
That portion is language work. And it's eating a disproportionate share of senior researcher time.
Why the PI ends up writing everything
The lab's research context exists almost entirely in the PI's head. The specific aims, the model organism, the funding history, the lab's established methods, the shorthand that means something specific in this lab but would need explanation outside it — none of that is written down in a form that's useful for rapid drafting.
When grant season arrives, the PI writes because the PI is the only one with enough context to produce anything that doesn't require a total overhaul. Postdocs and graduate students can contribute sections they personally worked on. Administrative staff can handle formatting and compliance. But the narrative binding — the framing, the significance statements, the integration across aims — flows through the PI.
This is a context architecture problem, not a writing problem. The PI isn't writing from scratch because they're the best writer in the lab. They're writing because they're the only person (or tool) with sufficient context to draft anything useful.
A language model that knows nothing about the lab produces generic academic prose. A language model that knows the lab produces a first draft the PI can actually edit instead of rewrite.
The documentation load beyond grants
Grants are the highest-stakes documentation, but not the only documentation that consumes researcher time:
- IRB submissions and amendments. Protocol narratives, risk assessments, consent form language — structured, formulaic, but requiring precise institutional and study-specific language.
- Progress reports. Annual and interim reports to funders that restate aims, summarize accomplishments, account for deviations, and preview upcoming work.
- Abstracts and lay summaries. Conference abstracts, public-facing summaries for institutional websites, descriptions for non-specialist audiences. Each requires a different register and level of technical detail.
- Lab protocols. Step-by-step procedural documentation for established methods — tedious to write, critical for reproducibility and onboarding new lab members.
- Internal communication. Lab meeting summaries, project status updates, correspondence with collaborators.
Each of these is language work. Each of them benefits enormously from context — knowing the lab, the project, the PI's preferred framing, the institutional requirements. And each of them currently represents a tax on the time of people who trained to do science, not administration.
Without lab context in place
- PI writes most narrative sections because no one else has sufficient context
- AI output is generic academic prose — correct but not specific
- Grant sections re-establish the same background every cycle
- Lab abbreviations, conventions, and framing must be re-explained each session
- Documentation takes longer than the science that produced it
With lab context loaded
- Claude knows the lab's research area, PI, model, and methods before drafting starts
- First-draft grant sections use correct specific aims language and established framing
- Progress report boilerplate generates in minutes, not hours
- Postdocs can produce useful drafts without PI re-explaining context
- Lab's institutional knowledge is in the file system, not just the PI's memory
What the context gap costs
A typical R01 application might require 60–100 hours of writing time across all sections. A meaningful portion of that — the background, significance, and innovation sections; the approach narrative for established aims; the progress report for renewals — is structured language applied to established facts. It's the kind of work where a strong first draft changes the time requirement from 8 hours to 2.
Multiply that by three or four annual submissions per lab. Multiply that across a department. The documentation overhead of research is enormous, and it falls disproportionately on the people whose time is most expensive and most scarce.
The specific aims page is the most-rewritten section of any grant. It's also the most templated — a genre with well-established conventions the funding community expects. A lab that has built its research narrative into a persistent context can generate a strong specific aims draft in under an hour. Most labs spend days on it because they're starting from scratch every cycle.
What this looks like when it works
A lab running well on this system has its research context written down and maintained: the PI's name and institutional affiliation, the lab's primary research focus, the model organism and experimental system, the key methods and equipment, the current funded projects and their aims, the lab's vocabulary (the abbreviations, the specific terms, the framing the PI uses when presenting the work).
When a postdoc starts drafting a conference abstract, Claude already knows the research context. When the lab manager generates a progress report section, the aims are already in the system. When grant season arrives, the first-pass significance and background sections take an afternoon instead of a week.
The PI's role shifts from primary drafter to editor and judgment-exerciser, which is the appropriate use of their time and expertise.
Field notes
I've worked with a research lab producing some of the most technically demanding science I've encountered. The language layer — the grants, the reports, the protocols, the documentation — is entirely separate from that technical work and entirely amenable to AI assistance. The barrier isn't capability. It's that nobody has taken the time to write down what the lab already knows in a form the tool can load. That's a few hours of work that changes every documentation cycle from then on.
R.P.
