Video summary

Never Hit Fable 5.1 Usage Limits Again

Main summary

Key takeaways

Technology

Main idea

The video explains how to avoid hitting Fable 5.1 weekly usage limits quickly in Claude Code by offloading “token-heavy grunt work” (file reading, edits, routine changes) to sub agents running cheaper models, while reserving Fable for high-level planning, orchestration, and complex direction.

Why the usage limit is a problem (Fable 5.1)

  • Fable 5.1 has a cap: you can only use up to half of your weekly limit on it.
  • If you reach the cap early, you must wait days for the reset or buy more credits.
  • Most people use Fable for tasks that don’t need it: the tasks are often not complicated, just token-heavy (e.g., reading files, making edits), so you end up paying “expert-level rates” for “intern-level work.”

Core feature: Claude Code sub agents to reduce Fable token burn

A sub agent is an additional AI Claude that can delegate a specific job to.

  • Claude sends it instructions; the sub agent runs the work in a separate conversation, then returns results to Claude.

Savings mechanism

  • Sub agents can run on different models (e.g., Sonnet or Haiku) that use far less usage than Fable.
  • Claude continues to “stay in charge” via Fable in the main conversation, but doesn’t perform every step itself.

Built-in/default sub agents (no setup needed)

Claude Code ships with two default sub agents:

  1. Explore agent

    • Acts as “eyes and ears” for finding project files.
    • Optimized for fast project discovery, reporting file locations so Fable doesn’t waste tokens searching.
    • Recommended tiny model: Haiku.

  2. General-purpose agent

    • Workhorse for tasks like research, building, or reviewing.
    • Can dynamically choose among models: Fable, Opus, Sonnet, or Haiku.
    • Useful when you need flexibility across task types.

Routing rule via claw.md (guide/tutorial setup)

To avoid rewriting instructions every session, the video recommends:

  • Add a routing rule to claw.md so Claude automatically divides work the same way.
  • The routing rule tells the lead agent (Fable) to provide sub agents with:
    • clear goal
    • scope
    • context
    • return format

“Real unlock”: custom sub agents (control thinking level)

The video argues default agents save tokens by changing models, but custom agents save more by controlling “thinking level,” which strongly affects cost.

  • Changing thinking level can improve benchmark scores, but:
    • costs rise much faster than the quality improvements.
  • Therefore, if a task can reliably be done at a lower thinking level, you shouldn’t pay for max thinking unnecessarily.

How custom sub agents work

  • Each custom agent is a Markdown file stored in:
    • ~/claude/.../agents/
  • The agent file components:
    • Description: tells Claude when to call it + which model/thinking profile to use
    • Tools: what it can do
  • Agents can even be color-coded for easier monitoring.

Example custom agent team (3 essential agents)

A specific workflow is described, with discrete responsibilities:

  1. Explore agent (built-in Sonnet 5)

    • Determines what’s going on in the project.
    • Reports to Fable so planning starts with context.

  2. Research agent (custom)

    • Checks documentation and outside sources.
    • Runs on Sonnet 5 Medium.
    • Has web search access and reads relevant pages.
    • Returns short answers with source links.
    • If it can’t verify something, it reports missing items to Fable.

  3. Builder agent (custom)

    • Actually implements changes by editing files and running tests.
    • Runs on Opus 5 at Medium thinking.
    • Receives a finalized brief from Fable and focuses on execution.

  4. Reviewer agent (custom)

    • Independently verifies what the code changes actually did versus the brief.
    • Runs on Opus 5 Medium.
    • Uses the original brief + actual changes to run relevant checks.
    • If issues are found, it sends back specific fixes, after which Fable accepts the result.

Division of labor summary

  • Sonnet: exploration + research
  • Opus: building + reviewing
  • Fable: planning + directing/orchestrating

When NOT to use sub agents (extra handoff cost)

  • Spinning up sub agents also costs tokens.
  • If the work is a minor change, letting Fable handle it directly can be cheaper than:
    • starting a sub agent
    • reviewing its output
  • Agents are best when they remove enough work from Fable to justify the delegation overhead.

Additional resource mentioned

The creator promotes a Founder OS boot camp and a “guide in the description” that supposedly includes:

  • exact system prompts for the three custom agents
  • copy-pastable setup instructions

Main speakers/sources

  • Speaker/Source: The video’s creator/instructor (promoting “Founder OS boot camp” and providing Claude Code agent setup guidance).
  • Primary system/tool referenced: Claude Code sub agents (default and custom), controlled via claw.md and agent Markdown files.

Original video