What problem does it solve?
Dynamic workflows and subagent stages in Claude Code often inherit the session model by default, leading to overspending on expensive models for mechanical tasks or underpowering stages that need deep reasoning. This Skill enforces explicit model and effort pinning on every stage and adds a cost gate before any workflow launches.
Core Features & Use Cases
- Per-Stage Model Routing: Applies a two-tier policy (opus at xhigh for coding and reasoning, sonnet at high for mechanical and support work) to every agent() stage, with haiku banned outright.
- Effort Calibration: Sets per-stage effort levels (low through xhigh) bounded by the chosen model, so trivial scans never burn high-effort tokens.
- Pre-Flight Cost Gate: Estimates agent count, model mix, and cost versus an all-opus baseline, then asks the user to Approve, Cancel, or Modify before launching.
- Script Verification: The /ultracost:check command flags stages missing a model pin, pins that mismatch the described work, and effort exceeding the model cap.
- Use Case: Before launching a 12-stage ultracode workflow, draft the script, run the estimate, see it costs 60% less than an all-opus baseline, and get explicit user approval before execution.
Quick Start
Ask Claude to draft a dynamic-workflow script with per-stage model and effort pins, then run /ultracost:check on the file to verify the pins and review the cost estimate before approving the launch.
