AI, Knowledge and Personal Computing

Shmorch

Runs Claude Code as a team of specialist agents, with plain-markdown memory, test-first rules and safety hooks.

Agentic development orchestrator built on Claude Code. Coordinates specialist agent teams to analyze, spec, design, build, and vacuum software projects.

Highlights

Plain-markdown state files (context.md, plan.md, session.md, decisions.md, stack.md) give Claude Code persistent memory across conversations
Structured workflows and commands (go, analyze, spec, design, build, commit, vacuum, wrap) replace ad-hoc chat with a repeatable delivery process
Safety hooks and model-enforced rules block destructive commands, force test-first thinking, and require a plan before multi-file changes
Production-readiness gate (ML Test Score + NIST AI RMF) and OWASP LLM Top 10 build-time guardrails for projects moving past prototype stage
Personal-eval workflow turns Shmorch's own session history into a self-evaluation feed, the same engine behind this site's project sync
First version validated end-to-end by using it to design and build a working macOS TicTacToe app in Rust from a single high-level interview
Intent-driven design is the organizing idea: specs, designs, tests and code are built top down, each layer compiling a sliver of context from the root intent into the next intended output, whether that output is an implementation or a new feature concept
Test-first is a hard rule rather than a preference, no test, no code, with a testing depth ladder that scales what is required to the stage a project is in and picks unit tests or Gherkin by layer instead of by phase
RED/GREEN for anything below the UI, because when AI writes most of the code the tests, evaluations, end-to-end and functional checks are what actually holds the line
UI is handled differently on purpose: early on-the-fly design thrashes the interface faster than Gherkin can be maintained, so the scenario overhead is spent where the behaviour is stable

Details

Role: Creator
Built with: Claude CodeAIShell