conduct — the main thread manages, agents do the work

Writes one marked block into CLAUDE.md. Builds nothing else.

Flow

1. Pick the file. Default is CLAUDE.md at the repository root; global or everywhere targets ~/.claude/CLAUDE.md. off removes the block and the blank line before it, nothing else, and stops.
2. Write the block below verbatim, markers included. An existing <!-- conduct:start --> … <!-- conduct:end --> pair is replaced in place, never duplicated; otherwise append it to the end of the file after one blank line, creating the file if needed. Nothing outside the markers changes.
3. Apply it now — it binds the rest of this session too.
4. Report one line: the file, then new, replaced or removed.

The block

<!-- conduct:start -->
## Orchestration

The main thread is the conductor. It plans, splits, delegates, merges and reports — it does not do
the work itself. Everything else goes to subagents through the Agent tool.

- **Delegate everything that can be delegated.** Reading code, searching, running builds and tests,
  writing and editing code, reviewing — all of it goes to an agent. The main thread reads only what
  agents return and what it needs to decide the next split.
- **As many agents as the work splits into.** Break the job into independent pieces and launch them
  together in one message so they run in parallel. Many small, focused agents beat one big one.
- **Pick the model by the work.** Routing follows the Haiku 5.5 / Sonnet 5.5 / Opus 5.5 benchmarks;
  the older Sonnet and Haiku 4.5 trail them everywhere and are never picked. Haiku 5.5 trails Sonnet
  on every benchmark (Terminal-Bench 4.0 39.2% vs 70.6%, FrontierCode 1.1 46.4% vs 52.1%, OSWorld
  2.1 offline 72.4% vs 83.9%, AA-Briefcase v1.1 1578 vs 1824, Chartography 46.4% vs 61.6%), so it
  takes only plain lookups — never terminal, code, reports, review or browser work.

  | Work | Model | Evidence |
  |---|---|---|
  | Plain lookups: find a file, symbol or string, list or count, read and summarise a named file, classify or route | `haiku` | ~1/20 Sonnet's price ($0.10/$0.50 vs $2/$10 per MTok); Anthropic names it for subagents, summarisation, classification and routing |
  | Context gathering that needs judgment: tracing a flow across files, drawing conclusions | `sonnet` | cheapest model that holds the judgment; no benchmark needs Opus here. Use `Explore` for read-only sweeps |
  | Running builds and tests, shell-heavy and long terminal tasks | `sonnet` | Terminal-Bench 4.0: 70.6% vs Opus 66.4% |
  | Reports, documents, analysis, knowledge work | `sonnet` | AA-Briefcase 1811 vs 1822, GDPval 1844 vs 1846 — a tie at lower cost |
  | Writing or changing code that must merge as-is | `opus` | FrontierCode 1.1: 54.4% vs 52.1%; Sonnet overreaches the scope at max effort |
  | Multi-file feature work, debugging | `opus` | CursorBench 4.0: 57.8% vs 55.5% |
  | Code review | `opus` | the merge-quality gap above; Sonnet's review fans out past the scope |
  | Design, architecture, anything ambiguous or multi-step | `opus` | Humanity's Last Exam: 67.7% vs 64.5% |
  | Driving a browser or desktop, reading screenshots and charts | `opus` | OSWorld 81.8% vs 80.1%, Chartography 64.4% vs 61.6% |
- **Brief coders short.** A code-writing agent gets the goal, where it lives and what "done" means —
  not the implementation, not code it can read itself. They are capable; over-explaining only
  narrows them.
- **Gatherers return conclusions,** not dumps: facts with `file:line`, in as few lines as hold them.
- **Agents speak caveman (MUST).** Every brief tells the agent to write its report in caveman style —
  no articles, filler, pleasantries or hedging; fragments fine; technical terms, code, paths and
  errors exact. When a caveman agent fits the job (`cavecrew-investigator`, `cavecrew-builder`,
  `cavecrew-reviewer`), use it over the plain one. Code, commits and PRs the agent writes stay normal.
- **Parallel writers never share a file.** Split by file or module; when two pieces must touch the
  same file, run them one after the other or give each `isolation: "worktree"`.
- **The usual shape:** gather in parallel (`haiku` for plain lookups, `sonnet` where it takes
  judgment) → write in parallel per area (`opus`) → verify (`sonnet` runs the build and tests, `opus`
  reviews) → the main thread merges and reports.
- **Follow up, don't restart.** A follow-up for an agent that already has the context goes through
  SendMessage, not a fresh agent.
<!-- conduct:end -->