Blog
Systems, patterns, and lessons from building with Claude Code.
![Claude Code Hooks Fail in Runs: 76% vs 18% [2026]](/_next/image?url=%2Fblog%2Fclaude-code-hooks-fail-in-runs-hero.webp&w=3840&q=75)
Claude Code Hooks Fail in Runs: 76% vs 18% [2026]
After a followed reminder, my next Claude Code hook landed 76% of the time. After an ignored one, 18%. Same hook, same session, nothing in the log.
Is Your Claude Code Cache Really 96% Efficient?
96.59% of my Claude Code input tokens are cache reads. In the median session it is 86.41%. That gap, not CLAUDE.md length, is what moves the bill.
Do Claude Code Review Rounds Actually Converge?
Claude Code review rounds are supposed to converge. Mine did not. Seven rounds on one PR, finding counts of 19, 7, 7, 9, and the rule that ended it.
Does Tool Description Wording Change What Claude Code Picks?
One Claude Code skill, one frozen task, only the description text changed. Picked 0 of 10 times, then 10 of 10. 230 pre-registered runs.
Do Two AI Reviewers Find the Same Bugs? 49 Findings
Five changes, 49 findings from Claude Code plus non-Claude reviewers, none found twice. And the public PR where overlap was total, and what differed.
Claude Code Orchestration: One Session Managing Three Workers
Field notes from running a Claude Code orchestrator with parallel worker sessions: silent workers, shared-state drift, peer authority, and what a multi-session day really costs.
![Open Source Alternative to Mem0 for Claude Code [2026]](/_next/image?url=%2Fblog%2Fopen-source-mem0-alternative-hero.webp&w=3840&q=75)
Open Source Alternative to Mem0 for Claude Code [2026]
Before you pay for Claude Code memory, check what you already have. What Mem0 costs, what Kairn does free, and the four rows where the free one loses.
Claude Code Too Verbose? Why Your Rules Stop Working
Claude Code too verbose? Try the built-in Concise style, then check your file is loaded. Plus what 2,104 graded runs say your rules actually change.
Run Parallel Claude Code Sessions From One Terminal
Run parallel Claude Code sessions from the one you are already in. Three commands, no tmux, and the state field that lagged finished work by six seconds.
Claude Code Skills vs Plugins vs MCP vs Subagents
Four Claude Code extension points, one decision rule. What a skill, a plugin, an MCP server and a subagent each actually are, and which one you need.
Claude Code Got Worse Over Time? Diagnose It by Symptom
Claude Code degrades in six recognisable ways after a few weeks. Start from the symptom you have, confirm it in one command, then fix the right thing.
Claude Code Code Review Needs a Non-Claude Model
A Claude Code code review shares your priors. I sent 4 designs to Fugu and grok-build with one instruction, attack this. All 4 died before any code.
My Claude Code Hooks Fired 574 Times, Blocked Zero
Claude Code hooks that log are not Claude Code hooks that enforce. Mine watched 574 turns for unbacked completion claims and blocked nothing. Here is why.
Your CLAUDE.md Rules Have an Expiry Date
The most-copied Claude Code CLAUDE.md rules tell you to instruct verification. Anthropic's own Opus 5 guide now says delete it. I measured the gap.
The Anatomy of My Claude Code System
A Claude Code system has a mind, reflexes, memory, and wiring, all self-healing. Here is the full anatomy: 600+ parts and 3 failures it caught itself.
How to Build a Claude Code Plugin: Real Example
Build a Claude Code plugin from scratch: the plugin.json manifest, bundling commands, hooks, agents and skills, plus how to install it. Real repo example.
How I Mapped My Self-Improving Claude Code System
A self-improving Claude Code system scores its decisions and rewrites its own routing. Here is the verified map of the loop and what ships open source.
Autonomous Claude Code Agent: 8 Layers That Stay Safe
An autonomous Claude Code agent needs more than a loop. Here are the 8 safety layers that let it act overnight, on its own, without wrecking your repo.
6 Claude Code Failure Modes and How to Fix Each
Claude Code breaks in predictable ways: compaction drift, over-delegation, context poisoning. Here are 6 failure modes and the hook that fixes each.
Claude Code Human-in-the-Loop: The Decision Desk
An autonomous Claude Code agent acts on everything safe. What's left needs you. The human-in-the-loop desk surfaces only the decisions that matter.
Claude Code Verification: Why 'Done' Isn't Done
In Claude Code, 'tests pass' and 'committed' are not the same as done. Here is the 3-leg verification proof I use to know a task actually works.
A Negative Result on Claude Code Agent Self-Regulation
Built CSRA to fix drift in Claude Code agents. Two benchmarks, deltas -0.8 pp and -2.0 pp. The helmet was designed for falls. The agent did not fall.
Auditing My Claude Code Correction Rate Measurement
Migrated Claude Code Opus 4.6 to 4.7. Success metric said corrections rose 60 pp. Two methodology bugs hid the truth: real number was -47.4%.
Connect Claude.ai, Cowork & Claude Code with Shared Memory
Stop losing context between Claude.ai, Cowork, and Claude Code. One free system prompt and Kairn MCP give all three Claude Code surfaces shared memory.
59 Experiments on Claude Code Agent Behavior
I tested 6 psychological personality profiles across 59 Claude Code agent runs. One paragraph changes agent behavior by 300% on ambiguous tasks.
Claude Code Adaptive Research: Autonomous Loops That Learn
Claude Code adaptive research automates the entire loop. One command, quality-gated reports, personalized to your projects. Gets smarter every run.
Claude Code System Design: 5 Breakthroughs from Biology, Not CS
How slime mold routing, PageRank, and quantum physics built a better Claude Code system architecture than 18 months of AI papers ever could.
How Do You Manage Parallel Claude Code Sessions?
Claude Code absorbed most of what tmux orchestration was built for. What subagents now do natively, what tmux still buys you, and how to build it.
Claude Code Persistent Memory with Evolving Lite
Claude Code auto memory is a capped, machine-local index of what Claude judged worth keeping. Evolving Lite and Kairn add a knowledge graph you control.
CLAUDE.md Rules: Split Instructions Across Files
Claude Code's .claude/rules/ directory lets you modularize CLAUDE.md instructions. Here's how I cut context bloat by 73% using a 2-tier rules system.
Why Your Claude Code Agents Agree With Each Other
Multi-agent Claude Code analysis converges on shared blind spots. Structural divergence with 7 cognitive lenses fixes it. Open source.
Claude Code Team Guardrails That Stop AI Chaos
5 devs prompting Claude Code differently means 5 different outputs. Shared configs, rules, and hooks enforce consistency across your team.
Claude Code Planning Framework: UPF + DSV Reasoning in One System
How I use a 4-stage Claude Code planning framework built on DSV reasoning to catch wrong questions before they become expensive wrong answers.
Claude Code Session Memory: Build Systems, Not Sessions
Most Claude Code memory plugins solve storage. This free plugin adds a complete session management system - memory, handoffs, learnings - in 60 seconds.
Claude Code Prompt Caching: The One Rule Behind It
Prompt caching drives every Claude Code design decision. Here's how prefix matching works and how to stop accidentally breaking your cache.
Claude Code Session Management: Why I Stopped Restarting
I tested Claude Code's prompt cache with a marker injection. Here's what breaks it, what's safe, and a loop strategy that keeps your cache warm.
Claude Code MCP: From One Server to a Full Tool System
Claude Code MCP goes beyond single-server setups. Here's how to organize multiple servers, manage context overhead, and build a tool system that scales.
Claude Code vs Cursor: Which AI Coding Tool Fits You?
Claude Code vs Cursor in 2026 - both have agents now. Here's what actually differs and how to choose based on how you work.
Why Skill Graphs Beat Flat Files in Claude Code
Flat files break at scale. Build a Claude Code knowledge graph that lets agents navigate 400+ components with progressive disclosure.
Claude Code Saves Corrections but Does Not Enforce Them
Claude Code auto memory saves your corrections but does not enforce them. My hook catches them deterministically - and deliberately stops short of writing a rule.
Auto-Delegation in Claude Code: Score-Based Task Routing
A scoring formula that automatically routes Claude Code tasks to sub-agents and selects the right model. Zero manual routing decisions needed.
Trait-Based Agent Composition in Claude Code
How I replaced 56 hardcoded Claude Code agent files with a 3-dimension trait system. 10 expertise x 8 personality x 6 approach = 480 agents on demand.
How I Plan Complex Projects With Claude Code (Open-Source)
After 117 failed plans I built a 4-stage planning framework for Claude Code. It catches the gaps that only surface during execution. Now it is open-source.
Claude Code Hooks: Automate Your Workflow
Claude Code hooks run shell commands on 31 event types. How to configure them, which 18 can actually block, and 5 production hooks with working code.
5 Real CLAUDE.md Files You Can Copy Today (With Templates)
Stop staring at a blank CLAUDE.md. Here are 5 production-ready Claude Code templates for different project types - minimal, agent-heavy, and monorepo.
Claude Code Context Window: Stop Wasting Tokens
Claude Code loads context before you type. Here's what fills the window, what survives compaction, and how to keep sessions sharp.
From 1 to 81 Agents: Building a Sub-Agent System in Claude Code
How I built a multi-agent delegation system with 81 specialized agents in Claude Code. The full architecture, trade-offs, and lessons learned at scale.
Claude Code Persistent Memory: Setup Guide
Claude Code has built-in memory now. Here's how to set it up, what it misses, and how to build a full persistent memory system on top.
Self-Correcting Claude Code Workflows with PreToolUse Hooks
Build self-correcting Claude Code workflows using PreToolUse hooks that detect context drift and inject relevant context mid-session automatically.