Reading · Peter's Context Design

Happy Friday!

It’s hot in Madrid and Belgium plays Spain tonight, so not much thinking this week.

I’ve started digging deeper into the idea of Context Design, to marry that with Context Engineering. AI (or agents) are only as smart as the context you give them, so there’s a lot there. More later!

Interesting this week 👀

After Forty Years, Still No Silver Bullet: Jorge Arango revisits Fred Brooks’s 1986 paper. Brooks’s argument was that the hard part of software was never coding, it was deciding precisely what to build, and named four properties (complexity, conformity, changeability, invisibility) no technology removes.

“As always, how to build gets easier. Knowing what to build doesn’t.”

The crisis of the what: Francisco Barrera Aros argues the line between AI handling the “how” and humans owning the “what” is already gone in daily work. Draws on Ethan Mollick’s “jagged frontier” research and a Scientific Reports study where LLMs fell for the Einstellung effect (fixating on the familiar solution) while human doctors didn’t, and defended wrong answers with full confidence.

“Judgment is built by getting it right and getting it wrong in real contexts, with consequences that land on you. Intuition, in a way, is scar tissue.”

How tech workers are feeling in 2026: a workforce splitting in two: Lenny’s annual survey of 5,920 tech workers. 49% feel “amplified” by AI, 19% feel “destabilized” or “diminished,” and that split predicts career optimism better than role, seniority, or company size combined (β = +0.39, stronger than the other three combined). Burnout jumped from 44.7% to 55.7% in a year. 82% say AI makes them more productive; the same people describe their thinking “rotting.”

“I don’t fully understand what I merge.”

Designers and researchers report the most AI anxiety, the worst-rated managers, and the lowest willingness to recommend their own field (NPS -49 for senior ICs vs. -5 for founders). Only 22% fear literally losing their job to AI; 51% fear being expected to do more for the same pay.

Fluent Enough to Pass: The Standard That Was Never Neutral: English proficiency tests (IELTS, TOEFL, PTE) can’t score real communicative competence (pragmatics, turn-taking, code-switching), so they optimize for what’s gradeable: five-paragraph structure, standard grammar, explicit meaning. A study found intensive IELTS coaching raised scores significantly without raising actual English ability. And the sting: researchers testing AI content detectors found they consistently misclassify non-native English as AI-generated, while native writing is correctly flagged as human.

How Standardisation Became Gatekeeping: Part 2. Cambridge’s first English exam ever given (1913, 3 candidates) failed all three, despite examiners’ own notes calling them fluent, because the test measured literary scholarship, not communication.

The revolution was the easy part: Dan Maccarone on the gap between the AI launch party and the decade of unglamorous integration work after. Cites the MIT stat that 95% of AI pilots fail (not the tech, a “learning gap”), and the mobile parallel: it took about a decade after the iPhone before mobile web traffic passed desktop.

“The companies that win the agent era make checking the work ten times faster, not doing the work ten times faster.” — Jakob Nielsen, quoted in the piece

Losing the spark in their eyes: Frota on what using AI daily actually costs: not memory, the moment of creating meaning in your own work. Cites research on how actors memorize lines, they look for meaning in the script, not repetition, as the parallel to design mastery.

Crafting AI Explanations for Every Role in Your Enterprise: NNGroup argues different enterprise roles (developers, sysadmins, domain experts) need different kinds of AI explainability.

Stop Reporting UX Activity and Report Business Outcomes: UX teams lose budget fights not because the work lacks value, but because they report activity and UX metrics instead of revenue, cost, risk, speed, and retention. Teams that can’t translate get filed as a cost center and cut.

The Discipline of Making Claims Data Can Hold: Saeideh Bakhshi’s framework for not overclaiming from data: name exactly what was measured, what comparison gives it meaning, and the strongest verb the evidence has actually earned (”is associated with” vs. “predicts” vs. “causes”).