INTD 318 · Guides
Setting up a project for an agent
You will have somewhere north of forty working sessions with an AI agent between September and December. Every one of them starts from nothing. This is how you build a project it can pick up cold — and how that same structure carries your venture from the first interview to the showcase without anyone losing the thread.
Why this exists
An agent has no memory of you. Not between sessions, and often not between one long conversation and the next. Everything it knows about your venture at 9pm on a Tuesday is what it can read right now.
Most people respond to this by re-explaining themselves. You paste in a bit of context, get a decent answer, close the tab, and do it again tomorrow slightly differently. Three weeks in, the agent has been told four incompatible versions of what your venture is, your outputs have drifted generic, and you cannot remember why you rejected the second idea.
The agent has no memory. Your project directory is its memory. Everything below follows from that one fact.
Which reframes what a project folder is for. It is not storage. It is the working context you hand a collaborator who is brilliant, fast, tireless, and has amnesia. Build it for that reader and you get two things for free: an agent that stays useful in December, and a documented process you are graded on anyway.
The shape
One directory. Numbered folders, because the numbers are the argument — you cannot honestly write 02_definition before 01_discovery has something in it.
Make this in week one, empty. An empty labelled folder is a question you have not answered yet, and you can see it. A folder you create in week nine is a folder whose contents you invented in week nine.
The three files that do the work
Everything else is evidence. These three are the ones that make an agent useful rather than merely available, and they are the ones people skip.
README.md — five lines, no more
What the venture is, who it is for, what stage it is at, what you are working on this week, and where to look for what. Five lines. If it grows past ten you have started writing a pitch, and a pitch is not orientation.
CONTEXT.md — the standing brief
The things that are true every session and that you should never have to retype: the customer, the problem in one sentence, the constraints you are working inside, your current riskiest assumption, and the vocabulary you use. If you have decided your users are “operators” not “customers”, that goes here, and the agent stops calling them customers.
DECISIONS.md — the highest-value file you will keep
Append-only. Dated. Every entry: what you chose, what you rejected, and why.
2026-10-19 — Priced per seat, not per usage.
Rejected usage pricing: four of six interviewees could not predicttheir own volume, so a usage price is unquotable to the buyer.Revisit if we land a pilot where procurement wants a cap.
The rejected half is the part that matters. Without it you will re-litigate the same decision in November, and the agent — which cannot see inside your head — will cheerfully help you rebuild something you already killed for good reasons.
This file is also most of your process book. Twelve weeks of dated decisions with reasons attached is the story of the project. Write it as you go and December is transcription rather than archaeology.
What lands where, week by week
The folders are not decorative. Each one fills during a specific stretch of the term and feeds a specific deliverable.
| Weeks | Folder | What accumulates | Feeds |
|---|---|---|---|
| W1–W2 | CONTEXT.md | Problem statement, first assumptions, the vocabulary | Value Proposition Canvas |
| W2–W5 | 01_discovery/ | Interview transcripts, market sizing with sources, competitor set, synthesis | Market Research & BMC · Project Proposal |
| W5–W6 | 02_definition/ | Value proposition, business model, the proposal itself | Project Proposal |
| W6–W9 | 03_build/ | Design tokens, components, the prototype, the build log | Design System & Prototyped Solution |
| W9 | 04_validation/ | Test sessions, observations, what changed because of them | Business Validation Report |
| W10–W11 | 05_business/ | Unit economics, pricing, go-to-market, onboarding | Business Validation Report |
| W11–W12 | 06_showcase/ | Process book, deck, the five-minute cut | Showcase + Process Book |
Nothing in the right-hand column is written from scratch. Every deliverable is an edit of material that is already in the folder — if the folder was kept. That is the entire argument for keeping it.
Session hygiene
Two habits, thirty seconds each, and they are the difference between a scaffold that works and a folder that rots.
Point the agent at README.md, CONTEXT.md and the one folder you are working in. Not the whole directory.
More context is not better context. An agent handed forty files will average across them; an agent handed the three that matter will be specific. This is the same discipline as a good brief.
Before you close the tab: did anything get decided? Append it to DECISIONS.md. Did anything get made? Save it into the folder it belongs to, with a name a stranger could read.
If you skip this, the session happened but the project did not move. This is the single most common failure and it looks like productivity while it is happening.
Naming
Dates first, lowercase, hyphens: 2026-10-14-interview-03.md. It sorts chronologically, it is unambiguous a month later, and it survives being pasted into anything.
What never goes in
Three categories. The first two will get you in real trouble.
- Participant identities. Interview transcripts get pseudonymised at the point of writing — P03, operations manager, mid-size clinic — not later. Names, contact details, recordings and consent forms live outside the project directory. You told people what you would do with their words; do that.
- Credentials. API keys, tokens, passwords. Not in a file, not in a comment, not "temporarily". A key pasted into a repo has been leaked, even a private one, because the history keeps it after you delete the line.
- Anything you would not show the panel. Not a rule about secrecy — a rule about honesty. If a number is a guess, label it a guess in the file. An agent cannot tell your estimate from your evidence unless you mark it, and it will reproduce your guess back to you in December with total confidence.
Scaling it
The structure above is right for twelve weeks and one person. Two things keep it working as it grows.
Ten-file rule
When any folder passes about ten files, give it its own README.md — three lines saying what is in there and which file to read first. An agent reads directory READMEs before it reads contents, so this is the cheapest navigation you will ever build.
One source per fact
Every number lives in exactly one file, and everything else links to it. The moment your market size exists in both the BMC and the pitch deck, they will disagree — and the wrong one will be the one you present.
Research to production, and the seams
The folders change kind three times over the term. The seams are where projects go wrong, so they are worth naming.
Evidence → argument. Around week five, discovery stops being collection and becomes a claim. The test: can you state your opportunity in one sentence and point at the file that supports it?
If the sentence needs three caveats, you have not synthesised — you have accumulated.
Argument → artifact. Around week seven, the work becomes a made thing. Your design system starts here, and it should be small: tokens, a type scale, colour, three components. Not more.
This is where an agent becomes genuinely powerful and genuinely dangerous — it will generate a great deal of plausible interface very fast. Direct it against your tokens, not against a vibe.
Artifact → evidence, again. Week nine, live testing. The prototype stops being the output and becomes an instrument for producing the next round of evidence.
The folder handles this fine. People do not — this is where prototypes get defended instead of tested.
The showcase is not a fourth phase. It is the first three, cut to five minutes. If the folder was kept, the cut is an editing problem. If it was not, the cut is a writing problem, in December, at the same time as everything else.
Where to go next
Once the scaffold exists, the next question is what the agent can reach beyond your files — and for anyone working with design files, how to get a Figma document into the directory without letting it become a second, competing source of truth.
INTD 318 Creating a Design Startup · Emily Carr University of Art + Design
Set in Space Grotesk and Archivo · The Overpunch System