claude-kit

Context Engineering in Practice

A system that makes AI development repeatable. Hooks enforce behavior. Skills map to design thinking phases. Agents evaluate project health. One kit, distributed across every project.

40+ hooks
10 phase skills
2 strategic agents
6 projects
The Problem

Two Problems, Not One

Inconsistent Output

Same prompt, different sessions, different implementations. Tailwind one day, raw CSS the next. Same design, different code. No standards.

Forgotten Corrections

You correct a mistake. The AI says "got it." Next session, same mistake. Every correction resets to zero. No memory.

The obvious problem is inconsistency. But the deeper problem is that corrections don't persist. You're training a model that forgets everything you taught it.

You Correct
New Session
Same Mistake
The Reset Loop
What I Learned

Three Versions to Get It Right

Each version solved a real problem. Each revealed the next layer missing.

v1

Workflows

  • Custom commands that read standards files
  • Context loaded at session start when it worked
  • No guarantee the files actually got read
Result: Unreliable. Sometimes loaded, sometimes didn't.
v2

Enforcement

  • Hooks intercept every tool call
  • Brain stores corrections across sessions
  • Awareness monitors for failure patterns
Result: Reliable. But no methodology.
v3 (current)

Design Thinking System

  • Skills map to design thinking phases
  • Agents evaluate project health
  • One kit distributes across all projects
Result: Enforcement + methodology + distribution

The insight: Enforcement handles HOW the AI should work. Design thinking handles WHAT to do and WHEN. Most people build one layer. You need both.

The System

Three Layers

Enforcement at the bottom. Methodology in the middle. Evaluation at the top.

Event-Driven Enforcement

40+ hooks across five categories: safety, tracking, context injection, quality, and lifecycle. Every tool call passes through them.

block-dangerous stops rm -rf and force pushes

enforce-specs blocks edits until you read the spec

mcp-security-scan catches secrets before they leave the machine

Design Thinking Workflow

Skills map to design thinking phases. Each is a method card with forcing functions, not a hardcoded script. GitHub issues capture decisions, not just tasks.

/research → understand the problem first

/build → commitment point: creates branch, starts work

/commit → docs check, commit, push, PR in one flow

Strategic Agents

Two agents that evaluate instead of execute. They run at key moments and create issues, not code.

Context agent (haiku): session start, evaluates project state

Phase evaluator (sonnet): after commits, checks project health

Both file GitHub issues when they spot gaps

Under the Hood

The Development Flow

Every task follows the same rhythm. Hooks enforce it at every step.

/research

Understand

/define

Scope

/ideate

Options

/build

Create

/test

Verify

/review

Quality

/commit

Ship

Merge

Done

Non-linear. Jump back when you learn something new. Skip ahead when the answer is obvious.

The Methodology

Design Thinking Applied to AI Tooling

Same methodology I use for everything. Understand the problem, build the smallest thing that tests the assumption, iterate.

01

I used AI for all my development work

Every project, every task. Real usage across six codebases, not a side experiment.

02

Workflows were unreliable

Custom commands that read standards files. Sometimes the context loaded, sometimes it didn't. No guarantee.

03

Enforcement solved reliability

Hooks intercept every action. Specs block edits until standards are read. The model can't skip steps.

04

But enforcement alone had no methodology

The AI knew HOW to follow rules but not WHAT to do or WHEN. I needed structure around the thinking, not just the execution.

05

Design thinking gave it structure

Skills map to phases. Agents evaluate project health. GitHub issues capture decisions. One kit distributes everywhere.

The kit is three iterations of the same question: what does AI need to be a reliable work partner? Instructions weren't it. Enforcement was necessary but not sufficient. The answer is enforcement plus methodology plus evaluation.

Interested in AI + Systematic Problem-Solving?

I'm exploring AI evaluation, product, and research roles. If you're working on something interesting, let's talk.

Or explore my other work