agentsclimarketplace

Claude cabinet

Skill gengshirong1128-boop/claude-cabinet

Stop paying the king to do clerk work. Cost-aware routing across Claude Fable, Opus, and Sonnet.

Install
npx -y skills add gengshirong1128-boop/claude-cabinet

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Cost-aware Claude model routing across Fable 5, Opus, and Sonnet. Use when a task needs Claude Code model selection, premium-model cost control, model delegation, explicit escalation gates, Fable/Opus review, Sonnet execution, or a reproducible routing and cost report.

SKILL.md

4.7 KB, as published. Nobody here has run it

Claude Cabinet

Stop paying the king to do clerk work.

Route each task to the cheapest Claude model that can complete it safely, then escalate only on evidence.

Cabinet Roles

  • Fable 5 / Prime Minister: Decide irreversible, high-blast-radius, strategically disputed work.
  • Opus / Reviewer: Resolve hard failures, architecture conflicts, security-sensitive work, and adversarial review.
  • Sonnet / Executor: Default for implementation, tests, scoped analysis, refactors with clear constraints, and routine work.

Do not treat model tiers as a leaderboard. Match authority and cost to the task.

Workflow

  1. Identify the task scope, reversibility, risk flags, failed attempts, and budget mode.
  2. Run scripts/route_task.py with a JSON input when deterministic routing is useful.
  3. Start with the recommended model and effort level.
  4. Escalate only when an explicit gate is triggered.
  5. After Fable or Opus produces a decision, downshift execution to Sonnet.
  6. Run scripts/estimate_cost.py with expected token usage before expensive workflows.
  7. Report the chosen model, reasons, escalation gates, cost assumptions, and first action.

Routing Rules

Start With Sonnet

Use Sonnet unless a stronger trigger applies. Sonnet handles:

  • scoped implementation and bug fixes
  • test writing and verification
  • routine code review
  • first-pass analysis
  • reversible work with clear acceptance criteria

Escalate To Opus

Use Opus when any of these are true:

  • two evidence-backed Sonnet attempts failed
  • security, auth, permissions, data migration, or data-loss risk exists
  • the change is architecture-level
  • evidence conflicts and cheap verification cannot resolve it
  • adversarial review is required before release

Escalate To Fable 5

Use Fable 5 when any of these are true:

  • the decision is difficult to reverse and has high blast radius
  • several strong routes remain strategically disputed
  • the objective function itself is unclear or contested
  • the user explicitly requests the highest-capability final verdict
  • long-horizon autonomous work exceeds a normal single-session task

Downshift

Once Fable or Opus provides a clear decision, constraints, and acceptance criteria, assign execution to Sonnet. Keep the premium model only for the next review checkpoint.

Execution Modes

  • live-delegation: The host supports model selection, subagents, or delegation tools. Invoke the assigned models and collect their returns.
  • handoff-only: The host cannot invoke other models. Produce copyable prompts and do not claim calls were made.

Always state the execution mode and invocation status.

Live Delegation Procedure

When the current model is Fable and the host exposes an Agent, Task, subagent, or delegation tool:

  1. Spawn a Sonnet executor with the bounded task, constraints, and acceptance criteria.
  2. Inspect Sonnet's evidence and verification report.
  3. Spawn an Opus reviewer only when an escalation gate triggers.
  4. Give Opus only the blocking evidence, risk, and decision question.
  5. Let Fable synthesize the returns and issue the final verdict.
  6. Delegate the approved bounded execution or follow-up back to Sonnet.

Record each actual invocation. If the tool does not allow model selection, switch to handoff-only.

Claude Code Notes

  • Prefer official aliases fable, opus, and sonnet unless the user requires pinned model IDs.
  • Use opusplan when Opus should plan and Sonnet should execute.
  • Verify model availability and current resolution before relying on an alias.
  • For third-party providers, pin model IDs with ANTHROPIC_DEFAULT_FABLE_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, and ANTHROPIC_DEFAULT_SONNET_MODEL.
  • Do not expose API keys or write provider configuration without explicit confirmation.

Read references/model-routing.md when configuring Claude Code aliases, provider pinning, or fallback behavior.

Output Contract

# ModelCabinetPlan

## Execution Mode
## Task Classification
## Cabinet Assignment
## Escalation Gates
## Downshift Plan
## Cost Assumptions
## Risks
## First Action

Include:

  • exactly one starting model
  • an effort level
  • evidence-based escalation conditions
  • a downshift plan
  • explicit cost assumptions
  • no claim of model invocation unless the host actually invoked it

Commands

python scripts/route_task.py examples/task.json
python scripts/estimate_cost.py examples/cost-route.json
python scripts/benchmark.py
python scripts/validate_package.py

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.