Dumping my notes from July’s liked tweets and GitHub stars. Big themes this month: deleting old prompt workarounds on newer models, turning browser .har captures into CLIs, and people finally sharing concrete dollar/time receipts from production agent runs.
Main takeaways from July
- Delete your old defensive system prompt hacks. Anthropic’s context engineering notes for Claude 5-era models [1] and Pi’s harness updates [2] point to the same thing: all those rigid scratchpad templates and
ALWAYS DO Xcaps-lock rules you wrote for 2024/2025 models actually hobble stronger models. Teams deleting 80%+ of their legacy prompt scaffolding saw better tool use and fewer weird loops. [3] - Mechanical checks belong in linters, not skills. OpenAI shared how adding repo-specific review rules in Codex bumped finding recovery from 58.3% to 98% on internal repos. Boris Cherny nailed the split: if a rule can be checked mechanically, put it in a lint script or CI check; save
SKILL.mdfor domain context and tribal knowledge. [4] - Capture the
.harfile once, then generate a CLI. Instead of letting a browser agent click around a web app on every run, do one manual logged-in session, export the.harnetwork trace, and have your coding agent build a typed CLI from those HTTP calls. Faster, cheaper, and way less flaky. [5] [6] - Have agents turn bug reports into failing tests first. Bun’s team runs an issue-repro loop that reads new GitHub issues and tries to write a minimal failing test before any fix agent touches the code. Even when the fix agent fails, a reproducible failing test saves maintainers a ton of triage time.
- Voice-ramble your messy intent before letting the agent code. Karpathy shared a great workflow for greenfield features: spend 10 minutes rambling on voice with the model and make it interview you until the spec is tight before it writes a single line of code. [7]
Real-world receipts & benchmarks
- Mitchell Hashimoto’s $45k invoice auditor: hooked an agent up to read-only email plus 3 years of construction invoices, quotes, and meeting notes to audit incoming 50-page scanned bills. It caught ~$45k in real billing errors, duplicate line items, and wrong-recipient charges that actually got refunded/credited. [8]
- Uber’s Agentic Pods: paired ~30 engineers directly with operations teams. City capital allocation across 150 cities dropped from 15 hours -> 30 minutes, financial pacing from 2 days -> 10 minutes, and marketing web QA from 2 weeks -> 50 minutes. [9]
- Rebuilding SQLite in Rust from an 835-page manual: Cursor ran agents on an 835-page spec to build a clean-room Rust SQLite replica passing 100% of their held-out test suite, with a 15x cost spread depending on how they mixed planner and worker models [10]. Good caveat from Charlie Marsh though: passing a finite test suite != surviving production SQLite concurrency and power-loss edge cases. [11]
Ops gotchas for overnight VPS agents
If you run unattended agents on a Hetzner or DigitalOcean box over Tailscale, watch out for Tailscale’s default 180-day SSH key expiry. @levelsio had an agent diagnose what looked like a production server outage via an independent rescue path, and it turned out the box was completely fine - Tailscale had just expired the node key and locked out SSH. [12]
Two things worth setting up on any box running background coding agents:
- Disable key expiry on your server nodes in the Tailscale admin console (and keep an out-of-band provider console login handy).
- Configure
zswap+ a real swapfile so a parallel Rust or TypeScript build doesn’t OOM-kill your agent harness at 3am.
Starred GitHub repos from July
- awslabs/aidlc-workflows - harness-neutral adaptive workflow steering rules (
/aidlc) that work across Claude Code, Codex, Cursor, Kiro, and opencode without locking your team into one vendor’s CLI - 9001/copyparty - single-file Python file server with resumable chunked uploads (
up2k), content deduplication, WebDAV, SFTP, and fast markdown/media browsing - exelban/stats - open-source macOS menu bar monitor for keeping an eye on RAM pressure, swap, and thermals when local models and Docker containers start fighting