The five layers of coding-agent context: Layer 1 — Task context (what to build now) Layer 2 — Session context (what was accumulated) Layer 3 — Repo instructions (what the repo knows) ← almost always…

The fix for Claude Code failures is in the repo, not in the prompt. Before your next session, check: → AGENTS.md exists and fits on one screen → It says what the project does in one sentence → It…

The new post covers: → coding-agents-need-better-repositories ← the full argument What goes wrong when repos don't tell agents what they need to know: - Architectural guessing (edits that break…

Install in 30 seconds: Works with Claude Code, Codex, Cursor, and any agent that reads markdown.

harness-experimental gives agents a repo-level harness: • AGENTS.md — what this project is and where to start • product/ — product contract, feature inventory, decision records • docs/TEST_MATRIX.md…

6/ This matters because a lot of agent failures are not model failures. They are context failures. The repo does not explain: - product intent - architecture boundaries - accepted risk - test…

This matters because agents compound when the repository teaches them how to work. A good harness helps Claude Code, Codex, Cursor, etc. answer: - what changed? - what can break? - how do I prove…

AI coding failures are often not model failures. The repo does not tell the agent: - what to read first - what constraints matter - what proof is required I'm building harness-experimental for…