Forgetting isn't the problem. Two dimensions of code quality get conflated. Individual pieces: AI nails those. Coherence across the whole codebase: AI is nearly incapable of it by default. Every task becomes a new file, a new abstraction, a new pattern. Six months in, your codebase looks like a house renovated by a different contractor for every room. Each is to spec but none of them fit together. Bigger context windows fix the first problem and don't touch the second. The fix is a blueprint the agent reads before it touches any code. Blog post

Opus 4.7's three thinking levers. Extended thinking with a fixed budget is gone. Opus 4.7 has three separate controls. Effort level (low→max, xhigh default) sets the model's thinking propensity. Adaptive thinking makes the per-step call on whether to think at all. Prompt nudges ("think step-by-step" / "respond quickly") override a single turn. Higher effort isn't a cap, it's a lean. That's why max is prone to overthinking on long agentic runs. Leave it on xhigh. Tweet · Docs

Don't trap your AI context inside one tool. Most founders store their context inside ChatGPT projects or Claude conversations. It works until you want to switch and then you're starting from zero. Markdown files on your computer fix this. Point any agent at the same folder and it has the same context instantly – Claude today, Gemini tomorrow, whatever's best next quarter. LinkedIn

Notes from France. I'm working from abroad this week and trying to actually take a break. The familiar feelings: guilt from resting when there's so much to do, plus the FOMO from an industry that moves this fast. I didn't make the progress I wanted on the Fidero website rebuild – new customer onboardings had to come first. Right call but annoying.

On my radar

Opus 4.7 is eating tokens. Max plan subscribers ($200/month) are hitting session limits for the first time. Two things help: drop to xhigh effort from max and stay under 300–400k context per session. If you're going to switch models to save tokens, keep Opus for planning and use Sonnet to execute, not the other way around. The PRD (product spec) is where reasoning matters most – skimping there is a false economy.

Claude Design. Anthropic launched their Figma competitor. Prompts to prototypes, full brand systems, animated decks. Figma stock dropped ~10% on the announcement. Early reality is mixed: token consumption is high (heard one person burned their weekly limit on a 7-slide deck) and the Canva MCP remains crap. Interesting direction, early days. VentureBeat