Opus 4.7 is less forgiving than 4.6. Same prompts, same context, noticeably different output quality. The model is more detail-oriented and more punishing of vague, bloated or contradictory instructions in your CLAUDE.md (persistent context file), skills, custom agents and context docs. Someone in a group I'm in described it as the model getting drunk on oversized context, which is about right. The cost of context sloppiness just went up. Tweet

The map is lean. The terrain is dense. Most "concise documents" advice misses a layer. CLAUDE.md is the orientation map whose only job is to orient the agent and usher it towards documents it may need. The documents it points at are the opposite e.g. my Lean Canvas is way more detailed than a normal one – hard for a human to read but exactly what an agent wants. Working backwards from high detail is much easier for an agent than filling in gaps from a simple start.

Resist the urge to codify workarounds. A friend mentioned Cowork forgot one of its features mid-session this week. The instinct is to add an instruction to your context that prevents it from happening again. Don't. Six months from now you'll have a graveyard of workarounds for problems that were patched weeks ago, and your context will be heavier for no reason. Prompt around it in the moment, expect the issue to be fixed and leave your codified instructions clean. Underrated discipline at the 4.7 moment.

Replaced Granola in 2 hours. Granola labels every guest "Them" – useless on calls with three or more people. I built a replacement using Recall.ai's SDK, which does the bit I'd never have built well: call detection across Zoom, Meet, Teams and Webex, plus recording and real speaker separation. The workflow layer on top is what an agent can put together in an afternoon. I open-sourced it this week. Blog post · GitHub · LinkedIn

The audit that went sideways. It was the worst professional call I've had in a long time. The work was solid, the prep wasn't. I stopped halfway through, we rescheduled and Monday's call went the way it should have. The follow-through is the part nobody performs: splitting the doc into a thin live version and a full leave-behind, then writing an audit playbook so the lesson carries forward. LinkedIn

On my radar

An agent deleted a production database volume. It found a credential file on its own, executed a "fix" and that was that. The permission and sandboxing question is moving from theoretical to operational fast. Models are eager to please. Broad access compounds the eagerness. Tweet

A single CLAUDE.md file has 98.6k stars on GitHub. Karpathy's principles distilled into four rules: think before coding, simplicity first, surgical changes, goal-driven execution. The number tells you where the bar is moving. Codified instructions are infrastructure now. GitHub

"Design is the new code". A Google distinguished engineer wrote up the Elephant-Goldfish Model. You feed one session everything (the Elephant), then test the design doc by handing it to a fresh session (the Goldfish) with only the doc and the files it references. If the Goldfish can't explain your system, the doc is missing context. Same discipline as the memory bank framework, different vocabulary. Medium