18 posts tagged Claude Code. ← Back to all posts
I measured the pace of my own GitHub output before and after AI entered the workflow — 8,803 commits across 123 repos. The inflection is hard to believe, and I have tried to believe it honestly.
The engineering behind PantryAtlas v0.2.0: an instant→refine→swaps architecture forced by the Raspberry Pi's embedding budget, a ranking formula that only ever settles downward, relaxed-JSON Gemma over grammar-constrained decoding, and offline-first sync that replays on reconnect.
How heat-metrics-lab Chapter 7 got two interactive surfaces — manual sliders and a live api.weather.gov fetch — wired to the same drift-gated metrics.js. Covers the playground skill's single-state-object pattern, the CSP-respecting ZIP-to-lat/lng pivot, a two-stage NWS fetch, and a solar-elevation estimator built from first principles.
Heat Metrics Lab is now public — nine chapters of scrollytelling about air temperature, heat index, and WBGT, each anchored on primary sources, with a JS↔Python drift gate holding 56× tolerance headroom, no analytics, deployed to Cloudflare Pages. The Anthropic-tooling counterweight to heat-protein-lab.
Two days after heat-protein-lab shipped on Google's stack (Antigravity 2.0, Stitch MCP, DeepMind Science Skills), I built the same class of artifact on the Anthropic stack (Claude Code, advisor, tufte-viz, subagent-driven-development, session-report). Both produced a 9-chapter scrollytelling explainer deployed to Cloudflare Pages. The session-report telemetry shows heat-metrics-lab cost 2.7× fewer tokens. The Anthropic side caught two pre-ship bugs that the Google side had no equivalent gate for. The Google side still has the clearest single advantage — Science Skills' structured domain data — and no Anthropic counterpart today.
Building heat-metrics-lab required getting two things right in Phase 1: a JS-vs-Python parity gate holding both implementations within 0.5 °F of known reference cases, and formula correctness against Liljegren 2008 and Rothfusz 1990. The drift gate solved parity. The advisor — a pre-implementation reviewer running a stronger model against the same conversation context — caught a 6.68 °F RMSE error in the WBGT globe-temperature formula before any implementer touched the code.
heat-metrics-lab Chapter 5 built a 1140-cell divergence map — and the tufte-viz skill caught a normalization choice that would have silently told the wrong story about WBGT vs heat index. A post about design-layer bugs that code review can't see.
Agents arrive at each session without persistent state. Externalization mechanisms differ in what they capture: intent, declaration, structure, execution. Treating them as one bucket obscures the architectural choice in each. This post separates the four artifacts produced by the work shipped in 2026.
How my planning stack across the OpenCastor ecosystem — CLAUDE.md, MEMORY, spec/plan docs, status files — kept hitting the same ceiling, and what changed when Dave Jeffery posted a prompt about a single-page HTML + JSON workflow visualizer. The result is workflow-atlas, a Claude Code plugin that sits next to CLAUDE.md as the graphical layer over the textual context Claude already reads.
A few hours after writing about workflow-atlas as a context layer for Claude, I shipped the execution-layer companion. mcp-tape records every MCP JSON-RPC call your agent makes; mcpreplay.dev lets you read what happened. Open source, MIT, alpha.
v1.4.0 ships three independent feature tracks that, together, change what registering a new piece of hardware feels like: hot-plug detection at the OS level, an announce/confirm UX over MCP that lets Claude talk the operator through the bind, and the first half of SP6's self-attestation loop — robots now sign their own spatial-intelligence scores with their RCAN apikey.
An introduction to the OpenCastor ecosystem — a single ROBOT.md manifest, runtimes in Python and TypeScript, an MCP server, a Claude Code plugin marketplace, a public registry with EU AI Act-aligned compliance endpoints, and the consumer-tier UX I've been focused on lately. Built almost entirely with Claude Code, on a SO-ARM101 arm I call bob.
A two-command install that wires your robot's ROBOT.md into any Claude Code session as structured MCP resources. Also: what RCAN and the Robot Registry Foundation do, and why this plugin belongs in Anthropic's hosted MCP catalog.
Between April 17 and April 21, 2026 I shipped ROBOT.md v0.1 through v0.8, an MCP server, a hardware pendant with ESP32-C6 firmware, and the landing page at robotmd.dev — across 364 commits in five repositories. This post documents the resource envelope (tokens, cost, cache-hit rate, commit cadence) and compares the observed working style to Thariq Shihipar's published recommendations for the 1M context window.
VS Code + Claude Code maintained context across weeks of SharePoint term store fixes through persistent memory files. Here's why decision routing matters more than token windows.
Vibefounding makes startups easier to start—and easier to kill. Here's what I learned from 11 days building Civqo and why the missing runtime layer matters.
Geoffrey Huntley's Ralph Wiggum technique for every platform. macOS, Windows, Linux, Raspberry Pi, and cloud servers. 10+ AI agents for autonomous coding.
182 commits after the kids went to bed. A cloud platform built with Claude Code and the Ralph Loop.