Agent Conduct

Agent Conduct

This is a meta-skill: a rule that applies across every stage rather than belonging to one. It covers two documents this workspace's own governance/README.md groups together as the AI-contributor-specific rules — ai-contributor-guide.md (the canonical owner for context-loading order) and agent-conduct.md (the canonical owner for contributor behavior rules) — read live from this plugin's own source files every time this skill runs, so there is no separate paraphrased copy anywhere for either to drift against. This is a distinct concern from the governance meta-skill: that one governs whether a document is canonically owned and acceptance-ready; this one governs how an agent behaves while working in this repository.

AI contributor guide — context-loading order (live)

!cat "${CLAUDE_PLUGIN_ROOT}/skills/agent-conduct/ai-contributor-guide.md"

Agent conduct — behavioral rules (live)

!cat "${CLAUDE_PLUGIN_ROOT}/skills/agent-conduct/agent-conduct.md"

Instructions

1. At the start of a task, load context in the order the AI Contributor Guide content above states — do not skip straight to a capability or feature document before the manifests and active process/stage definitions that resolve which one applies.
2. When a stage step names a role, resolve its competencies via the Execution Role Manifest as the guide above directs — execution roles are not product IAM roles and are not a new canonical owner of anything.
3. When blocked on a material decision that cannot be safely resolved from local context, ask a clarifying question through your environment's mechanism rather than guessing — per the Ask Questions section above. If a required output document is expected, record the question and its answer there, not only in conversation.
4. When a different execution role is assigned mid-session (with or without the conductor running the step), instantiate a fresh, isolated sub-agent for that role rather than continuing in the same context, per the Sub-agents Instance Per Role section above — this prevents state leak between roles' distinct permission boundaries.
5. Name the capability you need and resolve it through the Capability Model table above — never name a vendor in a persona, skill, or process definition. Climb the Search Strategy ladder in order and stop at the first rung that answers; run build/test/lint/container/version-control commands through the Execution capability.
6. Before writing anything new, apply Compose Before You Build and check Composable Skills above — another installed plugin may already own the rule. Treat every composed skill as optional, and never let one override a stop condition, a required output, or a gate decision.
7. Apply Lazy File Reading, Precision Code Edits, and Terse CLI Commands as your environment's tools allow — ranged reads over files longer than roughly 150 lines, targeted diffs instead of full-file rewrites, and limited-verbosity command output.
8. Point at a generated artifact, journal entry, or state-record path in conversational output instead of re-pasting or re-summarizing its content — per Durable State Pointers above; this applies to state records specifically and, more generally, to any large plan or content that already exists in a file.
9. If you are re-reading this skill after skills/agent-conduct/ai-contributor-guide.md or skills/agent-conduct/agent-conduct.md changed, trust what you just read above over any memory of a previous run — that live content is always the current source of truth for how an agent should load context and behave in this repository.