The reference doc that comes before any page: Before building a single page, an agent catalogues every component in the template – what it's for, what goes in it and the hard limits (a hero headline maxes at 10 words, a testimonial runs 25–45, etc). A contract, not a component list. The copywriter agent writes to those limits, so the copy drops straight into the build instead of getting reworked halfway through. Skip the doc and you just pay the time back re-prompting generic output into shape. Blog post

The ABZ order I hand every agent: A is where things are (status and project conventions). Z is where you're headed i.e. the bigger goal. B is what's next: the spec for the job in front of you. Ground the model in A and Z first, then give it B, and verify the result against both the spec and the conventions, not just the spec. A changes fast, so it gets rewritten after every B and folded into the PR (pull request). Hand over B alone and you'll wonder why the output keeps drifting. The memory bank framework is where A, B and Z live. Blog post

The spec didn't die, the reader changed: A PRD (product requirements doc) is still a product spec and you still need one. What changed is who it's for – the agent, not a human team. Same document, different reader, so you write it differently.

Agent setups degrade by default: The longer AI helps with your work, the more of its own context it creates – notes, instructions, whole documents it writes, then reads back next session. There are no guardrails on that growth, so it sprawls, and the sprawl is what quietly drags output down. Treat the context like data: give files owners, cap their size and prune regularly.

On my radar

The bill goes up while the price goes down: Cost per token keeps falling but I keep hearing the opposite: bills climbing and "Max" allowances disappearing, because agents just do more per request. Some also suspect their sessions get downgraded to a weaker model mid-task.

Models are worst at following their own plans: A long-read that ran different LLMs through a game of Civilization argued that the model best at writing a plan can be the worst at sticking to it once execution starts. Matches something I keep noticing. Article

AI detectors are noise: Someone ran two AI-assisted pieces through a detector and got "100% AI" for one and "100% human" for the other. Worth remembering next time one gets waved around as proof.