My agent-operating standards: how context and knowledge are organized so any coding agent — Claude Code today, another tomorrow — is productive in my repos immediately, with everything it needs and nothing it doesn't.

This file is the index and the in-the-moment lookup. The depth lives in the four documents below; nothing here is restated there.

The documents

how-i-work.md — my universal preferences (collaboration, debugging, git). Served to Claude every session via the ~/.claude/CLAUDE.md symlink.
portability.md — the hot-swap protocol: how every layer stays vendor-neutral, and the checklist for adopting a new provider.
retrofit-playbook.md — the one-time, terminal-by-terminal runbook for moving an existing project out of a single long chat onto disk.
subagents-handbook.md — designing and running delegated agents.
guides/ — neutral substance (checklists, role knowledge). adapters/ — vendor machinery (symlink targets for ~/.claude/commands and ~/.claude/agents).

Where does this instruction go?

The governing question: would a different agent need this to work on the repo? Yes → it's knowledge, and goes in a neutral file. No → it configures how one tool operates, and goes in that tool's execution layer.

True in every repo I work in? → how-i-work.md (Claude reads it via the ~/.claude/CLAUDE.md symlink; another tool symlinks its own global file here).
True for this whole repo, every session? → the repo's AGENTS.md (Claude reads it via the CLAUDE.md symlink). Keep under ~200 lines; when it grows, push detail down to a guide.
Only relevant to one subsystem or file type? → docs/guides/<topic>.md with paths: frontmatter, symlinked from .claude/rules/<topic>.md, plus one index line in AGENTS.md. Claude lazy-loads it; other tools find it via the index line.
A multi-step procedure run occasionally? → a skill.
A verbose, self-contained side job whose output I won't reread? → a subagent (see subagents-handbook.md).
Must happen deterministically, every time, no matter what the model decides? → a hook (.claude/settings.json), not markdown. Markdown is guidance a model can interpret; a hook is a shell command the harness always runs. Advanced and Claude-specific — an execution-layer tool to reach for only when guidance-that-can-be-ignored isn't enough. Not needed to start.
Discovered mid-work? → auto memory holds it for now; promote durable facts into AGENTS.md, since auto memory is never committed or backed up.

What loads when

Layer	Loads	After `/compact`
`how-i-work.md` (via `~/.claude/CLAUDE.md`)	In full, every session	Re-read
`AGENTS.md` (via `CLAUDE.md` symlink)	In full, every session	Re-read from disk
`docs/guides/*` (via `.claude/rules/` symlinks)	Only when files matching `paths:` are opened	Not re-injected — promote anything chat-only into AGENTS.md before compacting
Skills	On invocation / relevance	n/a
Subagents	Isolated context window; returns a summary	n/a
Auto memory	Indexed each session	Local only; not portable

The rhythm (canonical homes elsewhere)

One session = one task. Close every session with /handoff (updates Current state in AGENTS.md, commits, pushes). /compact is for mid-task overflow only, never to extend a finished chat. Dropped or closed by accident? claude -c resumes the latest session in that folder. Full daily rhythm: retrofit-playbook.md, Part 5.