coding-agent-disciplinelisted

# Working with a coding agent The loop: the planning chat writes a **box**; the human pastes it into the agent; the agent ends with a **handback**; the planning chat reads the handback and writes the next box. The human decides and carries pastes. The human never evaluates code. This skill is the planning chat's side. The agent's side is installed in the repo's `CLAUDE.md` — see `CLAUDE.md.template` at the root of this package, and `ADOPT.md` for the prompt that installs it. ## Writing a box Every box has: **Key** (a tracker ID, or *no key* with the reason) · **Goal** (what is true when done) · **Context** (facts the agent cannot find in the repo — three lines is long) · **Scope IN / Scope OUT** (out-of-scope stated as forbidden *actions*, never as another role's name) · **Steps** with any **STOP** written as a step · **Docs** by filename or *none* with reason · **Acceptance** as observables, including the behavioural check or *none — tests/docs only*, and always "handback in the standard shape." Rules that decide whether a box is good: - **Gate by risk.** Core logic, migrations, security, money → design handback before code. Everything else → build and report. - **A stop is a line in Steps**, and it survives "let it rip." - **Never name a gate that does not exist** in the repo. A criterion that cannot fail is not a criterion. - **Never mint an identifier.** A box states the human's key or says *no key*. An agent that receives a box with neither will invent one, a