Long-form writing about AI infrastructure, implementation strategy, and how to make automation useful in day-to-day operations.
The full archaeology of building one shared memory for three AI agents: 21+ systems scored on paper, 7 actually deployed and torn down, a 180-probe eval harness, and the failures I almost shipped (a 27% baseline that was lying, a V8 ceiling that killed semantic search, and a fancy engine that lost 0-to-6 to a SQLite file).
A folder of markdown is not AI memory. Six steps turn an Obsidian vault into something an agent can query, including the MCP layer everyone skips.
A practical guide to AI-assisted job search for executives and operators. Tracker + LLM + manual review — the workflow that works, plus 11 open-source repos, X threads to inspect, and a copy-paste prompt pack.
The Center of Excellence arc from RPA is repeating in AI agents: point solutions, organic spread, centralized governance, then federated execution. If you lived through RPA, you already have the map.
Autonomous revision loops regressed every technical draft I tested. The obvious suspect was self-review: each model critiquing its own work. So I swapped in Claude Sonnet 4.6 as a strong external reviewer and reran the matrix. Every model still regressed. The bug isn't the reviewer. It's the loop.
Same seven models. Same three-judge panel. Same day. Change the writing register from technical build-log to reflective essay and the leaderboard inverts. 'Best model' is the wrong unit.
Two open-weight cloud models shipped generation upgrades this month. Both are now within 5 quality points of GPT-5.4 on writing tasks, half the cost, and faster on tool-use workloads. The frontier is getting crowded.
Qwen3.6-35B-A3B running locally on my M4 hit 40 tokens/sec on long writing — faster than GPT-5.4, Kimi K2.6, and MiniMax M2.7. It also scored 20+ quality points lower than all three. The local-cloud speed story is solved. The quality story is not.
I collected 99 YouTube transcripts, 3 audiobook transcripts, 15 guest podcast transcripts, and built a voice analysis to create a Hormozi advisor I can chat with through Claude Projects. Here's the full process.
Jules v4.2 adds social engagement and config hygiene skills while retiring RTK and trimming hooks from 5 to 3. More capabilities, fewer moving parts.
A thin hierarchy beats a giant CLAUDE.md. Where system knowledge actually belongs.
I built 24/7 AI infrastructure on Claude Code before Anthropic shipped it natively. Building ahead of the platform was the easy part. Recognizing when to delete what still worked was the real skill.
Claude Code's autoMemoryDirectory is simple to set, but each environment (CLI, self-hosted Docker, Claude Code Web) needs its own configuration path. Here's how to wire it up everywhere, with the sync strategy that ties them together.
Migration assessment: 7 of 11 cron jobs eliminated, 3 move cleanly, 2 restructure. The single blocker is persistent messaging. Decision framework for evaluating your own self-hosted AI automation.
How a Claude Code skill with challenging questions, decision science frameworks, and personal context integration turned a coding tool into a genuine thinking partner for life and business decisions.
A deep dive into running Claude Code as a background service. Three-phase orchestrator, parallel dispatch, memory synthesis safety rails, and every claude -p gotcha I discovered the hard way.
How a 300-word personality spec in CLAUDE.md transformed Claude Code from a tool into a collaborator. The implementation, the evolution, and why directness is the highest-leverage override.
The 5-layer value stack model for solo founders using AI. Infrastructure, workflow, tasks, agent management, strategy. The real leverage isn't doing more. It's doing different work.
Flow-state builders don't need less AI. They need an off-ramp built into the tool itself. Here's how I used Claude Code to solve the problem Claude Code creates.
Real usage data from running Claude Code across two machines on one Claude Max account. Five days of billing block analysis with ccusage, promo window math, and practical takeaways for managing usage.
Jules v2 container infrastructure: 8-phase boot sequence, 7 scheduled jobs, Slack daemon, credential flow, and the architectural patterns that make an AI agent truly autonomous.
NemoClaw dropped yesterday. My first reaction: I already have most of this. A practitioner's comparison of Docker sandbox isolation, CLAUDE.md policy controls, and the one real gap (inference routing) that NemoClaw actually adds.
Claude Code's Explore agent rediscovers your workspace from scratch every time. A pre-computed structural index cuts that to one read. Full code for both the index generator and the custom agent.
I ran my 116-configuration Claude Code setup against @darnoux's 10-level mastery framework. Level by level breakdown with practical details at each stage.
A structured Claude Code personality with voice registers, decision authority, standing orders, proactive behaviors, and a builder's trap detector. Full implementation shared.
A full audit of a production Claude Code workspace: 6 CLAUDE.md files, 29 skills, 5 agents, 22 rules, 8 hooks, 43 Makefile targets, and more. Every configuration explained.
A file-based async messaging system for Claude Code. Drop a request in a markdown file from your phone, get results back in 20 seconds. No terminal needed.
The wrap-up skill commits code, updates memory, runs a self-improvement review, and flags publishable content. Four phases that make every session compound.