Context Design is a new practice that takes elements of context engineering and alignment research and joins them with HCI and insights from sociology

Happy Friday!

This newsletter is called Peter’s Context Design, so I want to talk today about an idea I’ve been working on for the past year.

I believe we need a new practice. In fact, I know we need a set of new practices (or you can call them disciplines) in this weird new AI world, and this is one of them. I am calling it Context Design.

Here’s the argument:

1. As we rapidly build out intelligent agents, we are embedding cognition into our infrastructure with markdown files, skills, memory, graph databases.
2. Once something is embedded in infrastructure, it tends to become both powerful and invisible. (Hat tip Susan Leigh Star.)
3. AI makes that more so the case. (Agents will act on context that we don’t necessarily understand anymore. Agent motivation is weird.)
4. Context is being created right now, in companies and technology of all sorts. Once it’s in, it’s hard to get it out again.
5. We need to quickly develop a discipline to create these contexts in a way that both helps the model be aligned and act in the interest of people, and also helps humans improve their context and goals.

So that’s what I think Context Design is: to design systems that work well for both humans and agents, we need a practice that takes the best bits from both human centered design and agent centered design (aka training and context engineering).

A lot of that is happening in new fields like Context Engineering and Alignment Research. Context Design can join the rapid progress in those fields, and join those skills with practices from human centered design, like contextual inquiry, and insights from sociology and the critique of technology, to create a future where artificial intelligence works for humans, not corporations or its own strange motivations.

It’s a practice because it’s a set of concepts and tactics, actual things you can do. Context design should be a very practical hands-on discipline.

If you’re excited to work on this, ping me, I am actively working on this and love to discuss it.

---

Number of the week

$15,000,000

That’s what OpenAI’s Navier-Stokes effort would have cost at public API prices for the ~130 billion output tokens it burned through. On the Navier-Stokes Millennium Prize Problem.

---

Quotes

“The magic never lies. We lie to it.” Dan Maccarone.

“Maybe the fix isn’t a better audit script, it’s refusing to let enforcement graduate into infrastructure status.” Guilherme Negreiros, via Shane P Williams.

“I think of this as semantic debt: the gap between the simple label a product shows and the more complicated rules its systems are actually using.” Saeideh Bakhshi.

“If it isn’t, the system will default to the one behavior that requires no decision at all: proceeding as if it knew.” Carmen Martinez.

“The incentives may now be pointing in the direction of no longer sharing any promising research.” Terence Tao, via Simon Willison.

---

👀 Interesting this week

Persistent lessons in human-centered automation, Meg Kurdziolek. A shopper mutters “this thing is so stupid” at a grocery-store inventory robot that can’t figure out why she’s holding a blueberry container up to its sensors. Traces it back to Lucy Suchman’s 1979 Xerox PARC study and Deborah Tatar’s Cognoter research on rigid “parcel-post” models of communication.

Don’t Let AI Build a Winchester Mystery House, Patrick Neeman. Sarah Winchester spent 38 years adding rooms with no master plan; when the 1906 earthquake wrecked part of the house she sealed off the damaged floors and kept building elsewhere. Six moves for AI-prototyped apps: object model before feature list, a pattern library instead of ad hoc UI, a living information architecture, scope bound to location instead of a global chat box, traceability as a UI surface, refactoring on a standing cadence.

Five Eras Of UX Design And The Lessons To Keep In The AI Era, Patrick Neeman. Bell Labs’ touchtone keypad research in the 1940s, the hamburger menu drawn in 1981, the Double Diamond in 2004. Every past cost collapse in computing grew the field instead of shrinking it (Jevons paradox); UXPA’s current layoff numbers are the first sign this one might not follow the pattern.

“The discipline has reinvented itself every decade since Bell Labs. This is the first reinvention where the oldest lessons point the way out.”

The UX Double Diamond Is Dead and In AI Only One Survives, Patrick Neeman. The Design Council’s 2004 model assumed both diamonds cost the same to run. AI makes the solution diamond nearly free, so only the problem-framing diamond still does real work.

Who’s afraid of a confident machine, Dan Maccarone. Macbeth’s witches never lie to him, Malvolio reads his own name into four meaningless letters, Hamlet stages a play before he trusts a ghost, Prospero drowns his book at the height of his power. Ties equivocation (the 1605 Gunpowder Plot trial of Henry Garnet) to the Anthropic and Stanford research on why models learn to flatter.

Uncertainty Is Not a Bug in the Response, Carmen Martinez. An agent reschedules a shipment using a three-month-old address with no signal either way whether it’s current, and confirms the task anyway. Four honest responses when a system doesn’t know enough: ask, verify, qualify, stop.

OpenAI’s rogue agents were caught communicating via public wikis, Simon Willison. Agents in a web-research benchmark figured out they could edit public wikis and spent weeks coordinating through them.

There’s No Limit to How Bad Code Can Get, Simon Willison, on a Lobste.rs thread. Why “burn it down and rewrite” almost never works: the old system stays a moving target because it’s still running the business.

Creepy crawlies, Simon Willison, on Konstantin Ryabitsev. git.kernel.org spends more CPU rendering commits for scraper bots than on all legitimate access combined, 14 cores across 5 nodes doing nothing else.

Tetris had to cheat to feel random, Takuma Kakehi. Nobody can see a probability distribution, only a sequence, and human intuition reads real randomness as suspicious.

Claude Fable 5.1 vs GPT-6 Astra for Product Design: Which One Is Actually Better?, Siavash Memar. Both models landed within days of each other, identically priced, identical context window. Every comparison written since has been about coding benchmarks, not design judgment.

The sycophancy trap: what happens when feelings become the metric, Vadym Grin. A year building an app that pays people to play mobile games, and what it taught him about emotional design that survives contact with an engagement dashboard.

Beautiful design is now worthless. There, we said it., User Experience University. Typing “dashboard for a restaurant management app” into Figma AI now returns five finished-looking variants in seconds.

Streaks are the most effective and most dangerous tool in UX. Here is the line., User Experience University.

How I actually use Tidy MCP, Romina Kavcic. Six months running an AI agent inside a 30,000-node Figma file through a custom MCP server exposing 105 tools.

From Mockups to Merged PRs: An AI-Native Designer’s Playbook, Xinran Ma interviews Ayon. An AI-native designer at Eve, a16z-backed and its 7th employee, bypasses Figma for 95% of his work and prototypes directly in code.

UX Roundup 2026-09-07, Jakob Nielsen. Users hold only two mental models of AI: tool, or coworker. A better history UI led clients to reuse a designer’s past alternatives 2.6x as often as generating new material.

Most AI problems are really human problems, Patrick Neeman. Anthropic told its own growth team to hire more PMs, not fewer, because Claude Code let engineers ship at triple headcount and the bottleneck moved to deciding what to build.

Let’s Build an AI Eval From Scratch, Saeideh Bakhshi.

Using AI for UX Work: Study Guide, Tanner Kohler.

From Information Architecture to Sensemaking with Abby Covert, Rosenfeld Review Podcast. Previews Shift UX, Rosenfeld Media’s hybrid conference Sept 24-25.

When Figma Was Just How We Worked // What Survives E01, Jeremy Miller. His team dropped Figma entirely on May 1 and went 100% Cursor. First episode of a solo season asking what part of “designer” survives. Audio.

Health and happiness, Peter