How we draft with agents, and how Elnora drafts for your lab

The number one thing everyone wants AI for, scientist or not, is everything around documents. Statements of work, reports, slide decks, figures, the whole pile. This is exactly how I draft mine at Elnora, and it is also the loop we built into Elnora itself as an AI agent that drafts lab reports from your data, because a scientist writing up a plate of data has the same problem I have writing a proposal: the facts are scattered across files, and the draft has to be right.

The loop I run

The foundation is a knowledge base. Ours is under a year old and already holds more than 2000 documents: call transcripts, previous applications, internal notes, every article worth keeping. It is just folders of markdown and CSV files, readable by a human in Obsidian and readable by an agent in the terminal. When I need a document, I tell the agent exactly where to look: pull the client details from the CRM folder, read yesterday's transcript, walk through the email thread, and draft me the statement of work as a new file in the vault. Our templates live next to the notes, so the agent knows the sections and what fills each one. It builds a task list, searches, and drops a first draft.

Then I read it paragraph by paragraph in the terminal, out loud, and the agent applies my edits into the file. After the words I run it past six specialised reviewer agents, contracts, IP, employment, corporate, compliance and risk, and decide which of their objections are real. After the review the agent renders the document as HTML in my browser using our brand file, iterates on what I say I do not like, and prints the PDF from the same browser. The email goes to my drafts folder and I press send myself. And everything that worked, the script, the template, the design decision, goes back into the knowledge base so the next document starts from it.

That is the whole method. Facts in one place, an agent that reads them, a human who reads the draft, and the result filed back where it came from.

The same loop, inside Elnora, for a scientist

A scientist's version of my statement of work is the write-up: the report after a screen, the deck for the group meeting, the workbook with the fits. The facts are just as scattered, across the plate map, the reader export, the protocol that was actually run, the paper that justified the design, and whatever the last person wrote down. So Elnora runs the same loop, and you do not have to set any of it up.

The knowledge base is the Elnora Knowledge Base: your protocols, readouts, papers and notes in folders the whole lab shares. Search takes a full question and returns the passages that answer it, each tied to the file it came from, so the agent, and you, can open the source in one step. When the agent produces something, it files the result into the folder it belongs to, which puts a result in front of the rest of the lab the day it is produced. That filing step is the one most labs skip, and it is the reason a knowledge base decays into a downloads folder.

The drafting is what I do with a statement of work, applied to data. Bind a plate map to the instrument's readout file and the analysis runs on one table keyed by condition. Ask for the dose-response fit and you get the curve, the potency value and the points behind it, with each condition compared against the control on the same plate and the excluded wells listed. Ask for the figure by name and it is saved under that name in your workspace. A conversation that ends in a deck or a workbook produces an editable file, built sheet by sheet with formulas that carry across sheets, and a saved style profile sets the colours, type and logos once for the group so every draft looks like it came from the same lab.

The review pass is built in. Three specialists share the work: a researcher for the literature, a data analyst for the results, and a critic that reviews the design and ranks each point it raises. Every answer closes with a source list naming the literature identifier behind each claim, so you can follow one statement back to the paper the way I follow a clause back to a transcript. I still read every paragraph of my documents myself. The critic does not replace that for a scientist either; it means the first draft you read has already been argued with.

And the last step, filing what worked, is where Elnora goes further than my setup. It carries your saved details, your own skills and your named protocol templates into every new conversation. Name a template and the next experiment is drafted to it. The knowledge base proposes its own tidying, which you approve or reject, so the folders stay usable without a person spending Friday afternoon on them.

What this is worth

I built my loop with a terminal, a folder of markdown, and a year of tuning. It works because the agent has the facts, a human reads the draft, and the result goes back into the pile. Most scientists will not build that, and they should not have to. Elnora is that loop with the science already wired in: the plate readers, the public databases, the fits, the figures, the review, and the filing. A write-up that took a week of copying numbers between files becomes a conversation, and the next one starts from what this one produced.