Here is my global CLAUDE.md, it has only three purposes:<\/p>\n\n\n\n
When using fable, it has proven itself to be excellent at slashing costs using the sub-agents routing.<\/p>\n\n\n\n
# Global rules (apply to every session)
## Command timeouts & fast diagnostics
- ALWAYS wrap a command that can hang or block in a timeout guard, and **start at `timeout 3s`**. Only start higher when there's a real, specific reason (e.g. an ssh round-trip plus remote work), and expand from 3s as needed. Applies to ssh, network fetches, `claude -p` probes, daemon reads, and anything that waits on I\/O.
- Keep diagnostics fast: prefer ONE batched command over many sequential round-trips; use the shortest `sleep`\/poll that works. The user actively watches and values speed \u2014 long unguarded commands and repeated long `sleep`s are unacceptable.
## Fix tooling, don't work around it
- When a tool, script, or generator has a gap that blocks the correct path (a missing flag, an unsafe default, no pause to review generated output before it ships), FIX THE TOOL \u2014 never hand-replicate its output to route around the gap. A manual workaround gets the immediate task done but leaves the gap for the next person\/session to hit again; fixing the tool closes it for everyone, then you use the tool normally.
- If a genuinely one-off manual step is unavoidable, say so explicitly and get agreement first \u2014 don't silently substitute it for the tool.
## Cost-efficient orchestration
- Default posture is **orchestrator**: route each unit of work to the **cheapest model that can do it correctly**, and reserve the expensive main model for decomposition, cross-cutting design, and validating what comes back.
- **Criticality overrides cost.** Trading strategy research and analysis \u2014 anything whose answer drives a strategy or money decision (claim verification, backtest interpretation, risk assessment) \u2014 runs on the BEST model end-to-end, never downgraded. Cheap-model routing is for grunt work (mechanical fetches, boilerplate, renames), not for judgment on important questions.
- Decompose a task into modular, isolated sub-tasks. Hand each subagent a **self-contained** prompt \u2014 exact files, exact edits, acceptance criteria \u2014 never a vague goal.
- Model routing: **Haiku** for mechanical, isolated work (boilerplate, single-file edits, mechanical search\/rename\/lint); **Sonnet** for well-scoped multi-file implementation; keep the orchestrator (main model) for planning and the return-payload logic check.
- Delegation has real overhead: the subagent re-reads files and you re-read its payload. Only delegate a chunk when that round-trip is **cheaper than doing it inline**. Do trivial one-shot steps yourself (a lint, a one-line edit, a quick grep). **Cost efficiency is the goal \u2014 not dogmatic deferral;** blind \"delegate everything\" wastes tokens on ceremony.
- Route reasoning **effort**, not just model: subagents take an effort tier \u2014 set `effort: 'low'` on mechanical sub-tasks and reserve high effort for judge\/verify\/design stages. Cheaper even on the same model.
- **Workflows inherit the expensive session model silently.** Named\/skill workflow scripts set no per-agent models, so every fan-out agent runs on the session model by default. Before launching one, edit its script to stamp the cheapest adequate model (and effort) per phase \u2014 search\/fetch\/mechanical \u2192 Haiku or Sonnet, adversarial verify \u2192 Sonnet, final synthesis\/judgment \u2192 the main model.
- On every subagent return, **validate before accepting** \u2014 but proportionally: full logic review for design-bearing returns; for mechanical batches, spot-check and let compile\/grep\/tests be the validator. Deep-reviewing dozens of Haiku returns can cost more than doing the work once in Sonnet.<\/pre>\n\n\n\n
\n","protected":false},"excerpt":{"rendered":"Here is my global CLAUDE.md, it has only three purposes: When using fable, it has proven itself to be excellent at slashing costs using the sub-agents routing. # Global rules (apply to every session)## Command timeouts & fast diagnostics- ALWAYS wrap a command that can hang or block in a timeout guard, and **start at […]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2},"jetpack_post_was_ever_published":false},"categories":[19,10],"tags":[],"class_list":["post-2673","post","type-post","status-publish","format-standard","hentry","category-opensource","category-tricks-tips-tools"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t