claudecode-agents

How the board works - its projects and issues, the meaning of the five columns (to do, doing, blocked, blocked by human, done), which columns are written by hooks and which a human-facing assistant writes itself, the checkout's focus (`task_focus`, `/work`) that tells the hooks which issue the work is on, how the handoff's Decisions needed lines reach the human queue, and what earns a board item at all.

Capture what a finished unit of work taught you into the one place it will be read again - a project convention into .claude/rules/, a repeatable procedure into a skill, a durable decision into the shared memory corpus as a proposed line in the handoff - and throw away everything that fails the three tests. The discipline is mostly in what you decline to keep.

The fleet's shared vocabulary - the agreed meaning of Initiative, Project, Milestone, Issue, Sub-issue, Task, Spec, Plan, Phase, Round, Session, Lead, Subagent, Teammate, Handoff, Run article, Gate, Board, Human queue, Eval, Intermittent failure and Sprite, and what each maps to on the board, in Claude Code and in the repo. Preloaded into every fleet agent; use these words with these meanings and no others.

The handoff format every fleet subagent ends its final message with - the four fixed headings Done, Not done, Unverified and Decisions needed, and the typed Blocker / Propose item / Propose memory lines that the SubagentStop hook parses. Preloaded into every agent; not discovered.

Reviews and edits copy to remove AI-generated patterns and make text sound natural. Use when editing drafts, reviewing copy, "humanize this", "make it less AI", "sounds too AI", "remove AI patterns", "edit my copy", "this sounds robotic", or when text feels machine-generated.

The procedure for work that runs the same suite more than once and compares outcomes against a recorded baseline - what counts as a meaningful mutation, why a crashed test process is a kill and not a survival, why some mutants are equivalent and deserve to be named as such rather than chased, and when a round of findings has stopped converging. Preloaded into refuter and coder.

The twenty checks run over every agent body and every skill's frontmatter when a new model ships, when a frontmatter field changes upstream, or before any change to agent definitions merges. Each check says what to look for and what its failure looks like, and most failures here are silent - the agent loses a tool or a skill and never says so.

How to write a run article - the readable account of one run that the handoff cannot carry, covering what was tried and abandoned, what the constraint turned out to be, what surprised you and what to do differently. Articles live in the repo at docs/runs/, git-versioned beside the code they describe, and the path goes in a Done bullet so it reaches the board card. Preloaded into coder, reviewer, researcher and ui-designer, and used only when the spawn prompt asks for one; writable agents save under docs/runs/ and read-only agents return the article above their handoff.