A doctoral program generates paper faster than it generates results. Meeting notes, ethics forms, grant files, course readings, three streams of literature, scripts, drafts, and the running list of things promised to people. The research itself is the smaller part of the pile. For my first year the pile lived in folders I searched by memory, and memory was the bottleneck.

What changed was not a new tool for thinking. It was an agent with read access to the whole vault and a set of rules about how to keep it. This post is about the organizational work the agent does: what it consolidates, what it checks, what it prepares. The research questions, the data and the people involved stay out of it, here and everywhere else on this site.

One vault, one map

Everything lives in one folder tree: a dashboard, the research streams, meetings, ethics, funding, milestones, conferences, reviews, coursework, admin, resources, templates, and an archive. The dashboard holds the hub notes: a home page that points at every active pathway, a progress tracker, a to-do list, a timeline, and a daily note per working day.

The agent’s first rule for the vault is to start every search from the home page rather than from the file system. That sounds trivial. It is the difference between an agent that reads the current plan and one that finds a stale note from a year ago and acts on it. The second rule is that status lives in dated notes and the hubs, never in the stable instruction files, so the agent always knows which document is allowed to be out of date.

The agent keeps the hubs current. When a meeting produces three action items, they land in the to-do with a date and a link to the meeting note. When a milestone moves, the timeline moves with it. I used to do this on Sunday evenings and often did not.

Meeting notes I write, consolidation the agent does

After every meeting I write notes by hand, in a template: who was there by role, what was discussed, what was decided, what I owe and by when. They are rough and they are mine.

The agent’s job starts after that. It reads the new note alongside the previous ones for the same group, and produces three things. A consolidated view of where that line of work stands, with the open questions carried forward from earlier meetings so nothing is quietly dropped. A short list of commitments, mine and others’, added to the to-do. And, when the meeting settled something, a proposed entry for the decision log, which I accept or edit.

Two rules shape this. Notes for different audiences are kept apart: the full note for my own record, and a separate copy of anything that was shared, filed under sent copies, so I can always see exactly what someone else has seen. And the agent never sends a note to anyone. It prepares; I send.

The decision log

Every design or execution ruling in a study goes into one file, under a dated heading, with a sequential number, the ruling quoted in my own words, the rationale, and a pointer to the memo or analysis that supports it. It is the system of record. Status notes and briefs point to it rather than restating it.

The agent reads the log first. Before it answers why something was done a certain way, and before it reopens a stance, it reads the entry and the memo the entry cites. Most of the time the question dissolves. When it does not, the agent proposes a new entry with the evidence attached and I rule, usually in a sentence, and that sentence is what gets quoted.

The log turned out to be the piece that makes a forgetful tool workable on a long project. A fresh agent, or a fresh me after a month away, cannot quietly relitigate something that was settled with a reason attached. And when a methods chapter later needs to say why a threshold is where it is, the entry already says why and when.

An operating manual per experiment

Each study gets a skill file that works as an operating manual. It has a table of where every piece lives: the vault folder, the scripts, the repository, the platform, the secured folder that holds anything identifiable. It has the commands that work, with the flags that work, and a mark on each one that writes to production. And it has the standing rules.

The rules are the part that matters. No paid API call of any size without first stating what will run, how many times, what it will cost per provider, what has been spent, and then getting a yes from me in the current conversation. One approval covers one named batch. Nothing team-facing goes out without my word, each time. Certain people are contacted only by me, never by the agent. Participant-facing material never mentions things participants must not know. Identifiable originals stay in one folder; the agent works from the de-identified copies. Every write to the live platform leaves an audit note, and a reload of a document that a participant has already opened needs my word for that document.

Dry runs are free and run first. The agent can simulate a full pipeline with no database and no API, print the exact text a participant would see and the logs the real run would write, and only then ask for the paid or production step. After each session on the platform it takes a read-only snapshot of the database into a dated folder, so the state at any point can be reconstructed.

I did not design these rules in advance. Each one is a line added after something went wrong or nearly did: a balance that ran out mid-batch, a note that almost went to the wrong audience, a number that an agent computed on the wrong generation of data. The manual is the accumulated record of what the agent must not do again.

Preparing material for supervisors

Supervisors read in a specific format: 12 point serif body, 1.5 spacing, real tables, no stray Markdown. The agent does the conversion and the checking. A script turns a Markdown file into a Word document with the right styles; a second script opens the result and refuses it unless the font, size, spacing and table count are exactly right and the text contains no smart quotes, em-dashes or leftover artifacts. Before the agent says anything about how a document renders, it exports to PDF and reads the pages, because word counts are not page counts.

The package rule is sample first. I see one document and eyeball it before anything is forwarded. And the genre rule for results is plain: a results submission is one to two pages of numbers, tables, figures and a few observation sentences, with no options, recommendations or questions. The supervisor gives direction. The agent used to want to add a discussion section to everything, and the rule in the file is what stopped it.

Ethics paperwork as a record

Ethics applications accumulate: the original, the resubmission, the confirmations, the guidance that came back. The agent keeps an application record with the exact wording that was accepted, and a lessons-learned guide for the portal, so the next application starts from what worked. It also enforces one rule of tense: ethics text describes research as prospective, “will be coded,” never as completed, because a review board reads tense as a statement about what has already happened to participants. Custodial data transfers can be past tense. Meetings that reviewed finished work are left out.

What the agent does not do

It does not attend meetings, decide anything, or send anything. It reads what I wrote, keeps it consistent, carries forward what would otherwise be dropped, checks what can be checked mechanically, and asks before anything that costs money or reaches a person. The thinking is the same as before. What changed is that nothing falls through the cracks between one week and the next, and the record of why things are the way they are is written down in one place, in my own words, with the evidence next to it.