MyClaudeCode
A mini coding agent used as an instrument to put numbers on unmeasured coding-agent advice.
1 · Purpose
MyClaudeCode builds a small coding agent to understand how tools like Claude Code work from the inside. Official guidance already explains how to set these tools up, but it does not measure the mechanisms it describes. This project runs experiments to find the numbers: when instruction files get too big, when subagents pay off, what compaction forgets, and where answer quality starts to fall. It is for engineers using coding agents on large codebases who want evidence, not folklore.
2 · Business benefits
3 · Status and how to join
Good first tasks this week:
- Run the existing usage tools over your own sessions and share the baseline split.
- Write the trace format for the mini agent's first version.
- Add one checked entry to the reuse ledger.
4 · Requirements
What it must do
- Provide a small coding agent with a trace from the first version.
- Split the always-loaded context baseline by layer.
- Force compaction on demand and record what kind of information it removes.
- Measure the break-even point where a subagent saves more than it costs.
- Measure how many skills can be loaded before selection breaks.
- Package a context-governance kit: status line, warning hook and pre-compaction backup.
How well it must do it
- Every published claim is backed by a number or a reproduced failure; unproven claims are cut.
- Measure context mechanics, which do not depend on the model; label anything model-dependent.
- Every experiment runs from a fresh clone by a stranger.
- Free open-weight models by default.
5 · Architecture, very high level
6 · Tech stack, per platform
| Component | Open stack | Azure | AWS |
|---|---|---|---|
| Agent loop | Hand-written Python, OpenAI-compatible client on Groq | Azure AI Foundry models | Amazon Bedrock models |
| Tools and editor | Python | — | — |
| Context assembler | Python | — | — |
| Transcript probes | Python over local transcripts, plus ccusage and cc-analyzer | — | — |
| Experiment runner | Python scripts | — | — |
| Governance kit | Claude Code status line and hooks | — | — |
7 · What needs to be done
Reposition and design 3 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Check whether the planned guide already exists | done | S | research |
| Decisions, glossary, reuse ledger and experiment list | done | M | architecture |
| Setup advisor and droppable skill spec | done | S | technical writing |
Measure the baseline 0 of 2 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Run existing usage tools over past sessions and read off baselines | todo | S | analysis |
| Baseline splitter and compaction forensics probes | todo | M | Python |
Mini agent 0 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Loop with a trace | todo | M | Python |
| Tools, invariants and a deliberately naive editor | todo | M | Python |
| Context assembly calibrated against the real tool | todo | M | LLM engineering |
Experiments and kit 0 of 2 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Compaction taxonomy and subagent break-even | todo | L | experiment design |
| Quality degradation threshold: cliff or slide | todo | L | experiment design |
8 · Roles
| Role | How many | Hours / week | What they do |
|---|---|---|---|
| Agent builder | 1 | 3-5 | Builds the mini agent and probes. |
| Experiment designer | 1 | 2-3 | Designs repeatable experiments and checks the numbers. |
9 · What you'll learn
- Why most coding-agent features exist because the context window is finite.
- How compaction works and what it can lose.
- When delegating to a subagent saves tokens and when it wastes them.
- How to check whether something exists before building it.
Level-up questions
- At what size does an instruction file start losing compliance?
- Is quality loss near a full context a cliff or a slow slide?
- Can the key experiments be automated without paid model calls?
10 · Links
- Links come once the work is public.