Blog
Just Published
I had Claude write a RAG eval harness. Here's how I made it auditable.
New I had Claude write nearly all of a two-layer eval harness for a client's RAG app. The design calls are mine, and they're what make it trustworthy: the judge model's own statement is stored next to the verdict my code derives, so a bad grade contradicts itself somewhere I can see. What I measured, what I rejected on the numbers, and where the numbers still can't help me.
Featured
Supergraphic Panel: chat-arch's design system, reusable in a prompt
Three URLs, one CLAUDE.md snippet, one prompt: how to pull chat-arch's dark chromed design system into any Claude Code project. With three recipes and the five places an LLM will miss.
design
How to Build Multi-Agent Systems with Claude Code
Subagents, skills, CLAUDE.md orchestration, headless execution, and state management patterns for Claude Code multi-agent systems. 50 claims verified against official Anthropic docs.
agentsclaude-code
Case Study
Case Study: Building a Self-Improving AI Agent That Runs an Etsy Shop
How I built ShopForge, a production AI agent with three-tier memory, Bayesian confidence scoring, and skill evolution, and what 10 days of real Etsy operations taught me about self-improving systems.
case-studyagents
Series
Browse by topic
2026
- I had Claude write a RAG eval harness. Here's how I made it auditable.I had Claude write nearly all of a two-layer eval harness for a client's RAG app. The design calls are mine, and they're what make it trustworthy: the judge model's own statement is stored next to the verdict my code derives, so a bad grade contradicts itself somewhere I can see. What I measured, what I rejected on the numbers, and where the numbers still can't help me.evalsagentsclaude-code
- Chip Huyen's AI Engineering, Applied: Grading My Production Systems Against the BookEvery idea in Chip Huyen's AI Engineering, graded against my production LLM systems, plus where the book and current Anthropic guidance genuinely pull apart.book-notes
- Command's LLM wiki: a self-updating memory layer that has to prove what it writesI built Karpathy's LLM wiki pattern into Command, my Claude Code operator: 134 provenance-tagged pages, a grep gate that blocks fabricated quotes, a nightly digest pipeline that was reset after a cheap model wrote a wrong verdict into it, and a human-gated publish wall.agentsclaude-codetooling
- Claude Code deletes your old session logs after 30 days by defaultThe cleanupPeriodDays default is 30 days, so Claude Code deletes session transcripts older than that at startup. Set it higher in ~/.claude/settings.json to stop the deletions. What the setting covers, what was already gone on my machine, and why raising it can't recover the past.claude-codetooling
- Command: the design decisions behind an operator that must prove its fixesThe key design decisions in Command, my autonomous Claude Code operator: why the watcher is deterministic and free, why it authors per-repo config instead of loading it, why every fix carries a pre-registered success criterion, and what I want to see from it next.agentsclaude-codeevals
- Parse the spec as claims: the CI drift guard behind an agent-readable design systemProse design specs drift the moment nobody's looking, and coding agents now read them as ground truth. The two-stage CI guard in chat-arch treats every palette-table row as a falsifiable claim: what it catches, the contrast bug that earned its own token, and the geometry drift it sailed right past.designclaude-codeagents
- Adversarial Subagent Teams: A Retrospective71 invocations of the adversarial-subagent-team pattern in Claude Code and Cowork, verified against the cited PRs, commits, and arXiv references. Across six headline claims: three confirmed clean, one confirmed with caveats, one rejected on falsification, one not supported by the data.agentsclaude-codecowork
- chat-arch Next: Site Survey, and the Project Idea You Forgot You HadClaude conversations split across claude.ai, Cowork, and Claude Code CLI fragment your real work-streams. Site Survey is the planned fix in chat-arch.chat-archtoolsai-workflow
- Supergraphic Panel: chat-arch's design system, reusable in a promptThree URLs, one CLAUDE.md snippet, one prompt: how to pull chat-arch's dark chromed design system into any Claude Code project. With three recipes and the five places an LLM will miss.design
- Alex Hormozi's $100M Offers Applied: Packaging Expertise as a One-Person Technical ShopEvery concept from Alex Hormozi's $100M Offers ranked by leverage for an independent engineer who already gives away deep expertise but hasn't packaged it into a real offer. Honest assessment with action plan.book-notes
- Two Patterns That Changed How I Think About Multi-Agent SystemsAnthropic's latest harness research reveals two underappreciated patterns: why agents can't evaluate their own work (and what to do about it), and why you should be actively removing complexity from your agent harnesses as models improve.agentsclaude-code
- Alex Hormozi's $100M Leads Applied: One Engineer's Implementation PlanEvery concept from Alex Hormozi's $100M Leads ranked by leverage for an independent engineer with a blog, no email list, and zero lead magnets. Prioritized action plan with dependencies and research gaps.book-notes
- Everything Broke at Once: Triaging Claude Desktop After a Week of OutagesFixing three Claude Desktop failures on Windows: 'VM service not running' in Cowork, Google Drive connector returning empty results, and persistent 'This isn't working right now' errors, with PowerShell diagnostics, log analysis, and root causes.claude-codebug-fix
- Fixing 'VM Service Not Running' in Claude Desktop Cowork on WindowsThe same error message covers at least six different root causes. Here's the diagnostic decision tree with PowerShell commands, log locations, and fixes for each one.coworkbug-fix
- From Audible Library to Interactive Bookshelf with AIHow the Claude browser extension, Claude Code, and the Google Books API turned 292 audiobooks into an interactive reading page with covers, a pie chart, and two view modes.site-build
- How to Build Multi-Agent Systems with Claude CodeSubagents, skills, CLAUDE.md orchestration, headless execution, and state management patterns for Claude Code multi-agent systems. 50 claims verified against official Anthropic docs.agentsclaude-code
- How I Built an MCP Server Starter Kit in a DayEvery MCP server I've built starts with the same scaffolding: HTTP client, Pydantic validation, dual response formats, 70+ tests. So I extracted the pattern into a reusable starter kit that turns any REST API into an MCP server for Claude, Cursor, and other AI assistants.mcp
- Where Claude Desktop Actually Stores Your Cowork ConversationsThe session manifests don't contain your conversations. The community tools don't know where to look. Here's the actual path to every Cowork message, tool call, and response, plus an audit log with per-session costs that nobody knew existed.cowork
- Fix: Cowork's file_upload Tool Returns 'Not Allowed' on Every OriginThe Claude in Chrome file_upload tool fails with CDP error -32000 on every site. The fix is a single Chrome extension setting. Here's the one-line solution, the diagnostic that found it, and the wrong assumptions along the way.coworkbug-fix
- Case Study: Building a Self-Improving AI Agent That Runs an Etsy ShopHow I built ShopForge, a production AI agent with three-tier memory, Bayesian confidence scoring, and skill evolution, and what 10 days of real Etsy operations taught me about self-improving systems.case-studyagents
- Fix: Claude Desktop Freezes on Startup (Cowork Scheduled Task Bug)High-frequency Cowork scheduled tasks can brick Claude Desktop on Windows and macOS. Root cause: the catch-up mechanism. Fix: rename one JSON file.coworkbug-fix
- Reverse-Engineering Claude Desktop's Cowork Session FormatThere's no official schema for Claude Desktop's Cowork session files. So I read 146 of them, diffed the fields, and documented what I found, including some surprises about what Anthropic stores locally.cowork
- How I Actually Build Software with Claude CodeAfter 1,500+ commits of AI-paired development, my workflow has narrowed to three tools with distinct roles. Here's what stuck, what didn't, and a real example of GPT catching a security bug in a PR.claude-codemeta
- Contributing to the MCP Python SDK: Exposing progress_callback in ServerSessionHow I identified a gap in the Model Context Protocol Python SDK, implemented a fix, navigated CI failures, and got a PR ready for review, all in a single day of AI-paired development.mcp
No posts match this filter yet.