What a week! A 5 day Design Sprint at Oxford University's Future of Humanity Institute, with AI Safety researchers from @FHIOxford and @CSERCambridge, designing a strategic role-playing game for decision makers about the effects of AI. With @anderssandberg@rossgritz
The agent that wrote your spec is the worst possible reviewer for it.
It remembers too much.
Google just open-sourced a fix: the Elephant-Goldfish Model (EGM).
โ Elephant: holds your project context, drafts the spec
โ Goldfish: reviews it with zero memory of the conversation, just the doc + the code
โ You: resolve trade-offs, approve, decide what gets built
Every assumption you agreed on in chat but never wrote down is invisible to the Goldfish. So it trips on the gaps your coding agent would hit later. Before a single line of code.
โ 7 skills, incl. brainstorm, draft, review & drift report
โ 5-stage design review
โ Reviewer rules that ban flattery
โ Works in Claude Code, Codex, Antigravity, ChatGPT & Cowork
โ Apache 2.0, 20K+ stars
Fastest test: run /egm:egm-review on a design doc you already have. ~60 seconds.
Beginner's mind, packaged as a plugin.
Search "google elephant-goldfish-model" on GitHub.
Your coding agent isn't running out of intelligence. It's running out of room.
One Playwright snapshot = 56 KB of context. 30 minutes in, 40% of the window is gone. Then /compact wipes what the agent was doing.
Context Mode fixes that:
โ 315 KB of tool output becomes 5.4 KB (98% less)โ Sessions go from ~30 min to ~3 hoursโ Agent writes a script & reads only the answer: 47 file reads (700 KB) become 3.6 KBโ Edits, tasks & decisions survive /compact via local SQLite + BM25โ 17 platforms: Claude Code, Cursor, Codex, Copilot, Gemini CLI
31K+ stars. Free & source-available.
Think in code NOT in context.
On GitHub: mksglu/context-mode
First answer in a customer interview is pablum (Bob Moesta's word). Shallow, not wrong.
Keep asking why. Once the real friction surfaces, people enjoy talking about it.
Most interviews stop one question too early.
Your coding agent ships a feature in 20 minutes. Then you spend an hour clicking through it.
e2e is an open source testing framework built for that gap:
โ Write the goal in plain English: agent.act('upgrade the workspace to Pro')โ Assert in plain English, or with normal locators, in the same testโ Verified agent steps get recorded and replayed with zero model calls until the UI changesโ Web (Chromium, Firefox, WebKit) + iOS & Androidโ Bring your own subscription, API key, or local modelโ Docs ship in node_modules so your coding agent can read them
Apache 2.0. 9K+ stars.
Goals NOT selectors.
npx e2e initOn GitHub: tester-army/e2e
We stayed up till 3am wiring real data into a prototype.
Real data = every edge case. We spent the night making it work instead of thinking about what we were testing.
If you're testing an idea, don't make it real.
Find the human moment in every workflow and make it thin.
Agent fills the queue. Human approves, revises, or rejects. Agent executes. Humans own judgment and taste, the agent owns everything around it.
Recording + playbook on the Evermuse blog.
Ian Pilon used to hand field research to engineers and wait.
Now he leaves the customer's site, prototypes that night, and goes back the next morning.
The rate of learning is daily. That's the real shift.
Something I took from studying Buddhism: thought and feeling depend on each other.
AI mimics the steps of thinking remarkably well. Ask it for the post about what you believe that others don't, and it can't write it. It doesn't believe it.
That gap is where judgment lives.
Every AI product I look at has a memory story. Most are too generic to matter for product.
Remembering yesterday's chat is table stakes.
What product teams need: patterns in customer behavior, usage tied to revenue, and whether the loudest voice is one person or a real trend.
You can spot AI-built UI from across the room.
Inter everywhere. Purple-to-blue gradients. Cards nested in cards. A rounded icon tile above every heading.
Every model trained on the same SaaS templates.
Impeccable gives your coding agent taste:
โ 24 design commands (polish, audit, critique, distill, harden...)
โ 61 slop detectors that run with no LLM, no API key
โ A hook that checks every UI edit while the agent works
โ PRODUCT.md first, so design starts from audience & purpose
Claude Code, Cursor, Codex, Copilot + 13 more. Free & open source. 76.8K stars.
Taste NOT templates.
On GitHub: pbakaus/impeccable
Ian Pilon told my favorite Crowdcast story:
Every stakeholder wanted their product above the fold.
Then the team saw the brokers' brochure drawers. Top ten: worn. Bottom ones: dusty.
The users had already done the information architecture.
Nobody argues with dust.
The pull toward autocratic leaders is really a pull toward 'let's start over.'
But there's no undemocratic country better to live in than a democracy.
Keep the democracy. Restart everything AI touches instead. That's where big change belongs.