INTD 317W02ADD / DROP DEADLINE TODAY

Context
engineering.

THE SPINE SESSIONTUE 15 SEP 2026

Last week the course demonstrated that context matters. It never taught how to build one. That is this session, and the studio output is literally two of their four notebook files.

Add/drop is today — expect roster movement. Do not ask anyone to commit to anything; the domain claim is deliberately next week.

Prep check: live client repo open in a second window, six-part handout printed, two concept-video/product pairs queued, one paper with a strong limitations section.

Workflow Autopsy.

TEN MINUTES
Two of you. Show us how you actually worked this week — screen, files, mess and all.
WHAT HELPED
One thing that saved you time or made the work better.
WHAT WASTED AN HOUR
This is the more useful half. Nobody is being marked on being efficient.

Next week's volunteers before you leave today.

First one of the term, so it sets the norm. If they present a tidy fiction, ask one specific question about the mess — "where did the file actually live?" — and the honesty level for the rest of term is set right here.

INPUT · 45 MIN

Setting up
the work.

Ten beats in 45 minutes. The two that must not be cut are beat 04 (the live repo) and beat 06 (iterate the instructions, not the output). If you are behind at 9:15, compress 07 through 09 into one slide and keep the scanning beat short.

The mental model.

The model has no memory of your project, no access to it, and no stake in it.

Everything it knows about your work, you put there. Everything you leave out, it invents — smoothly, in the same confident register as the parts it got right.

That is not a flaw to be fixed with better phrasing. It is the operating condition.

Resist the urge to anthropomorphise here, and resist the urge to over-correct into "it's just autocomplete." Both are shortcuts that stop them thinking. The useful frame is the practical one: what is in the room when you ask.

Anatomy of a working context.

GoalWhat are we actually doing, and for whom
ConstraintsWhat is fixed, what is off-limits, what cannot move
ReferencesExamples of good. And examples of bad.
Non-goalsWhat we are deliberately not doing
House rulesThe mistakes not to repeat
StateWhere the work actually is right now

Six parts. Most people supply one, and it is usually a compressed version of the first.

Hand out the one-pager here, not at the end. They should be holding it while you talk through the table.

Push on references, both kinds. Examples of bad are the part everyone skips and the part that does the most work — "not like this" is far more legible than any adjective.

Persistent, or per-task?

PERSISTENT

Lives in a file that loads every time. Who the project is for, the constraints, the house rules, the standing non-goals. Written once, corrected as you learn.

PER-TASK

Belongs in the message. What you are doing right now, the specific material in front of you, the one-off constraint.

Most people put everything in the message and rebuild the whole situation from scratch every session. That is the single most expensive habit in this room, and it is invisible because each individual instance only costs two minutes.

A real one.

LIVE CLIENT PROJECT

Six memory files, maintained across months. A file that refuses to start work until it knows which client. A rules file where every entry exists because something broke once.

WHY THIS AND NOT A DIAGRAM

It has a commit history. You can see when each rule was added and what went wrong the week before it appeared.

Live window · switch to the repo

Ten minutes, live, not screenshots. This is the most credible ten minutes available in the course and no other instructor at the school can run it.

Scroll the history. Open one rule and tell the story of the failure that created it. That single move teaches beat 06 before you get to beat 06.

Fallback if you cannot get online: screenshots of the file tree, one rules file, and one history panel. Capture them the night before.

Examples beat description.

Two examples of the output you want do more than a paragraph describing it. Three do more than a page.

This is not a fact about models. Hand a paragraph of adjectives to a junior designer and you get the same result: something that satisfies every word you wrote and none of what you meant.

BEAT 06

If you correct the same
thing twice, that's a
context bug, not an
output bug. →

The line the session hangs on. It is the difference between someone who prompts and someone who builds a system.

Make it concrete immediately: "you fixed the tone in message four and again in message eleven. The second fix was free to make and cost you the whole thread. Put it upstream once and it stops happening for the rest of the project."

Know when to start fresh.

WHAT ACCUMULATES
Contradictions, abandoned directions, decisions you reversed, three versions of the same constraint.
WHAT IT COSTS
The model has no way to know which version you meant, so it averages them, and the output gets quietly worse in a way that is hard to attribute.
THE MOVE
A clean thread with good context beats a long thread with drift. Take what you learned, put it in the file, start again.

Two things people get wrong.

ACCESS VS DESCRIPTION

There is a difference between telling a model about your data and letting it read your data. They produce very different work and very different risks. One sentence for now; we do this properly in week eight.

CONTEXT IS CURATED

Not dumped. Stale decisions, contradictions, and everything-you-happen-to-have all make output worse. What you leave out is a design decision.

Scanning, compressed.

WHERE IT SURFACES
Preprints, patents, teardowns, conference demos, and the occasional very good blog post by someone who actually built it.
ALWAYS READ THE LIMITATIONS
Every serious paper has one. It is the most honest section and the least read. It tells you what the demo did not show you.
RENDERINGS LIE
A concept video is a design artifact about a product that does not exist. Treat it as a claim, exactly like last week.

Apple, Knowledge Navigator, 1987 · via DigiBarn / Internet Archive · CC BY-NC-SA 3.0

Show the two concept-video/product pairs fast — thirty seconds each. The gap does the teaching. Then hold up the paper with the good limitations section and read two sentences of it aloud.

The file is local and plays with the wifi off — that is deliberate, and worth thirty seconds of your own credibility to point out. Space and the arrow keys belong to the player while it has focus, so scrubbing will not jump the slide. Click once on the video first.

Pair one, and the only one fully rights-clean: Apple's Knowledge Navigator, 1987. Play about thirty seconds — the agent in the bow tie, the professor asking it to pull papers. Then ask what shipped. Almost all of it, thirty-eight years later, except the part the video is actually about: that it understood him. The concept video promised comprehension and delivered retrieval.

Pair two: Magic Leap, reusing last week's demo against MIT Tech Review's 2018 hands-on. ⚠️ Not committed — the only copy is a YouTube reupload, so it needs the network or an advance download. Do not build the beat on it.

Break.
Ten minutes.

BACK AT 9:40HAVE YOUR NOTEBOOK REPO OPEN WHEN YOU SIT DOWN
STUDIO · 60 MIN

Build yours.

This is not an exercise.

What you write in the next hour isbrief.md and patterns.md — two of the four files in your graded notebook.

You do not have a domain yet, so write it for the project you think you might do. You will revise it after week three. Revision is visible in the history and revision is the point.

Say the "not an exercise" line explicitly. It is what stops the notebook reading as documentation homework for the rest of term.

Forty minutes. Six parts.

brief.md — WRITE IT FOR A COLLABORATOR JOINING ON MONDAY
GOAL          what we are doing, for whom, and why it matters
CONSTRAINTS   what is fixed. money, time, bodies, law, physics
REFERENCES    two things that are good. two things that are bad.
NON-GOALS     what we are deliberately not doing
HOUSE RULES   starts empty. it fills up as you get things wrong.
STATE         where the work actually is today

If a section is honestly empty, write “not known yet.” That is real information. Inventing a constraint to fill a heading is worse than leaving it blank.

Walk the room. The most common failure is a goal written as an aspiration — "improve the experience of X." Push for a deficiency: what specifically is bad, for whom, and how would you know it got better.

Now run your first scan through it.

COLD FIRST
Ask for recent developments in your area with no context at all. Save it.
THEN WITH THE BRIEF
Same question, your brief.md supplied first. Save that too.
THEN THE HARD PART
Take three claims from the better output and try to find the primary source for each.

Every claim goes into
research.md traced to a
source, or marked
[unverified]. →

THE TRACEABILITY RULESTARTS NOW · ENFORCED ALL TERM

State the consequence plainly: an untraced claim in a graded notebook is treated like an uncited quote. Not as a threat — as a definition, so nobody can say they did not know.

[unverified] is a completely acceptable answer and carries no penalty. Say that clearly or they will quietly drop the claims they could not source, which is the opposite of what you want.

Cold against context. Your own material this time.

TWO MINUTES EACH
Read one line from the cold output and one line from the context output. Do not summarise them — read them.
THEN ONE SENTENCE
What did the second one do that the first one couldn't?
AND THE HONEST CASE
If nothing much changed, say so. Thin brief, thin gain. That is a finding about your brief and it is worth more than a polite improvement.

Not everyone — sample six or seven and keep it moving. Note who had a genuinely thin brief; they are the ones to check on during week three.

Next week you commit.

DOMAIN CLAIM
Care, labour, food, mobility, memory, dying, play, shelter, money, language, energy, waste. Or argue for another. Two per domain, first come. One no-questions swap at week five.
COME WITH TWO
A first choice and a second. Think about it this week rather than deciding in the room.
RESEARCH PACK
Set next week, due week five. Nothing to do yet.
AUTOPSY
Two names for next week before you go.
NOTEBOOK
Commit brief.md and patterns.md tonight, even if they are rough. The history matters more than the polish.

Reiterate that the domain is thirteen weeks and that the alternate-format options from week one apply to everything, so nobody should pick a safe domain to avoid exposure.

Fix it upstream
and it stops
happening. →

W03 · TUE 22 SEPPRIMARY RESEARCH, AND WHAT AI CAN'T DO FOR YOU

Speaker notes

next
back
scroll next & back
swipe next & back
s speaker notes
o overview & jump
t start the 8:30 clock
f fullscreen
esc close

Turn your phone sideways

These slides are a fixed 16:9 canvas. In portrait they shrink to an unreadable strip; landscape gives you the whole slide.