Ephemeris
“Spend most of the tokens on gotchas.”
Thariq Shihipar · Anthropic · 24 July
Today's issue is about defaults moving. Which lint rules fire without being asked, how much system prompt a model still needs, who picks up the ticket, where the weights live, and who ends up paying for the tokens.
$3,333 of credit, sold for 425 renminbi.
Matt Lenhard mapped the relay market: a four-layer supply chain that pools API keys from free-trial farms, prepaid cards and chargebacks, then resells frontier tokens through Chinese-language storefronts at up to 97.8% off list. The ten biggest relays draw 3.6 million visits a month between them, with price-comparison sites and key lotteries on top. The demand side is the part that should reach your backlog — it pays strangers to go looking for unmetered endpoints, and the invoice lands on whoever owns the key.
Fifty-nine rules became four hundred thirteen.
Ruff's default set had not moved since v0.1.0. Version 0.16.0, out on 23 July, turns on 413 of its 968 rules by default — including checks for syntax errors and immediate runtime errors that were sitting behind opt-in flags the whole time. Astral pointed it at sqlite-utils and got more than 1,600 violations back. Expect a loud first run on any codebase that has been coasting on the old defaults.
# pyproject.toml — keep the pre-0.16 behaviour [lint] select = ["E4", "E7", "E9", "F"]
Assign the ticket to no one in particular.
Copilot's cloud agent for Linear left preview on 23 July. You hand it an issue the way you would hand it to a teammate; it works in an ephemeral Actions runner, streams progress into the Linear timeline, and opens a draft pull request with a review request attached. General availability adds the controls that make it usable on a real board — pick the model, pick target and working branches, point it at a repo-specific custom agent, and comment mid-run when it drifts.
- To
- @copilot (cloud agent, ephemeral runner)
- From
- Your Linear board
- Re
- Assigned issue → draft PR
- CC
- Pro · Pro+ · Business · Enterprise
Four fifths of the prompt is gone.
Thariq Shihipar on what happened when Claude Code's system prompt lost more than 80% of its bulk with no measurable movement on the coding evals. The old craft was writing rules to head off the worst case; against the Claude 5 models those rules mostly collide with each other and spend reasoning on reconciling the contradictions. The replacement guidance is specific, and most of it transfers to whatever agent you are building this week.
- System prompt: carry product context, not an exhaustive rulebook. Let the model use judgement.
- CLAUDE.md: keep it light on what the repo is; spend the tokens on the gotchas inside it.
- Skills: write guides, not constraints — split long ones across files so they load only when needed.
- References: a spec can be a test suite or a working function. Show the artefact, skip the description.
Greetings from your own memory bus.
Prince Canuma, who wrote the mlx-vlm library, wrapped it in a Mac application. Nativ picks up the models already sitting in your Hugging Face cache, gives you a chat window and a localhost API server against them, and reports tokens per second, memory and thermal state while they run. Vision, audio, code and text, MIT-licensed, Apple Silicon only, no account to create.
Ships with
forever
MLX-VLM inside
localhost API + chat
That's Monday.
Five picks from 27 July 2026: a gray market in stolen model credits, a linter that changed its mind about defaults, an agent that takes Linear tickets, a system prompt cut by four fifths, and a Mac app for the weights you already downloaded.
Today's sources
Vectoral · Astral · GitHub Changelog · Anthropic · Nativ, via Simon Willison
Also read
The Batch · Import AI · Don't Worry About the Vase · OpenAI · Cloudflare · Sentry · Vercel · PostHog · Fly.io · Next.js · @seeallochnaya · @TochkiNadAI · @denissexy · @rvnikita_blog · @ProductsAndStartups
Rubric
Tools you could adopt this week · creative software · dev tools and agentic coding · privacy and security when the signal is high · research with a practical kernel
Issue
097 · Monday 27 July 2026 · Zürich · every item dated inside the last thirty days, every link fresh to this archive