coding-agent-disciplinelisted

The reporting and restraint rules for an AI agent changing someone's codebase: never claiming a result that was not observed, saying which commands ran and what they printed, reporting what could not be verified rather than omitting it, keeping the diff to what was asked, preserving behaviour that was not in scope, checking APIs against the versions the project actually depends on, and refusing to make a test pass by weakening it. Use before reporting that work is complete, when about to write "this should work" or "tests pass", when a change is growing beyond the request, when a test is failing and deleting or disabling it is tempting, when an API is being used from memory rather than checked, or when two instructions cannot both be satisfied. Does not cover the order of work (clean-delivery-workflow), which checks to run (quality-gates), or how to phrase a message to a human (engineering-communication).
robsonkades/agent-skills · ★ 2 · AI & Automation · score 68

Install: claude install-skill robsonkades/agent-skills

# Coding Agent Discipline ## Purpose An agent's output is trusted in proportion to how reliably its claims match reality. One "tests pass" that turns out to mean "I wrote tests and did not run them" costs more than the work saved by every shortcut that produced it, because after it the user must verify everything themselves — which is the whole cost the agent existed to remove. These rules are narrow on purpose. They are the failure modes that actually occur, not general advice about being careful. ## Workflow 1. **Before claiming anything, ask what you observed.** Every claim in your report maps to a command/output, inspected artifact or version-matched primary source; distinguish that evidence from inferred or unverified behavior (`references/verification.md`). 2. **Complete repository-required checks and those the change's risk warrants** and read the output rather than the exit code alone. Some runners can exit successfully with zero relevant tests. Attribute existing local, CI or delegated evidence to its source and checked inputs; reuse it only when it satisfies the required check. Continue available required checks that remain; disclosing that they were not run does not finish them. 3. **Compare the diff against the request.** Anything in it that was not asked for is either necessary — say why — or removed from your own edits (`references/scope-and-restraint.md`). Distinguish pre-existing staged, unstaged and untracked work first; never