[ STATUS: INTERCEPTING_SIGNAL_STREAM ]
SYSTEM
ENTROPY
LOGS
Notes on what I'm building — AI tooling, systems, and what breaks along the way.
Delete-First: The Ablation Habit
The people getting the least out of Claude Code have a 400-line CLAUDE.md and add a rule every time it slips. The people getting the most delete. Here's the ablation workflow, with a real before/after.
ALL_FRAGMENTS
11 ENTRIES LOGGEDAlgebra for Agents: Orchestrating the Swarm
Sequential, parallel, fan-out. A small set of operations that compose into work one agent can't do. Here's the real fan-out audit script, the referee that keeps it honest, and loops vs routines.
Unhobbling: Goals, Not Steps
You take a model that can plan, read a whole repo, and run your tests, and you cram it into a single-line edit. The horse can run — you tied its legs and you're proud it can still hop. Here's how to cut the rope.
Verification: The One That Actually Matters
Most prompt advice is cope. The single habit that moved my success rate more than any trick: give the model a mirror. Tests, types, screenshot diffs, and hooks that won't let it call broken code done.
Anatomy of an Eve Agent: Five Files and Nothing Else
An agent is a directory. A model file, a markdown loop, a typed schema, a folder of verbs, and a test suite that says whether it still works. Here's one taken apart, then the build order that starts with the eval.
Eval-Gating: Don't Let an Agent Accept Its Own Output
A judge subagent with zero tools, a zod schema the runtime enforces, and an if statement on a boolean. Generate, grade, then accept, retry, or park — the closed loop that stops an agent grading its own homework in prose.
Eve Agents: Why Twelve Specialists Beat One God-Agent
I built a fleet of twelve agents instead of one that does everything. Not because twelve is better at the work — because when one of them breaks, I know which one. Here's the roster and the reasoning.
The Orchestrator Capstone: Composing Twelve Lessons Into One Agent
Agent #12 introduces nothing new. It's forced fan-out plus eval-gating plus durable state, wired together — and a scorecard that scores all twelve at once. Including the row where the capstone is red.
Root Lockdown: Delete the Tools and the Agent Has to Delegate
My research agent ignored its own instructions and did 40 web searches itself instead of fanning out. The fix wasn't a better prompt — it was taking the search tools away. Structural constraints beat written ones.
Using Claude Code Better: The Beginner's Guide
Most people use Claude Code like fancy autocomplete and leave 90% of it on the table. It's a terminal agent that reads your repo, runs your commands, and checks its own work. Here's how to actually drive it.
Using Claude Code Better: The Intermediate Guide
You already live in the terminal with Claude Code. This is the gap between it works and it works like a senior engineer next to you: ablation, real permissions, verification hooks, and orchestration you can paste.