kieiken/ultracode-token-optimization

This is a set of execution rules (skills) for token optimization that are loaded when writing scripts using orchestrations (agent() / parallel() / [pipeline() / [[T3], etc.) using Claude Code’s Ultracode (Workflow feature).

Background

Fable 5 (Orchestrator), × Codex (Executor), × Heterogeneous Reviewers. From the actual operations where multiple projects ran side by side, we extracted successful and failed patterns and reused them as Claude Code’s Skill.

The optimization target is not the “token unit price” itself, but redo (heavy industry) such as reverts→ re-revisions→ and re-reviews**. We have clearly defined patterns that easily lead to rework, such as crossing roles boundaries, leaking raw logs, completing checks only for self-reporting, and granting decision-making rights to reviewers, and we make them a practice that must be applied before writing scripts.

Contents

• SKILL.md — The main skill itself. frontmatter (trigger keyword) + the following contents:
• Actual constraints of the Ultracode Workflow runtime (agentType list, model inheritance behavior, budget actual specification of mechanisms, etc.)
• Model layout tables by role (orchestrator / executor / reviewer / mechanical tasks)
• Policy for replacing ‘heterogeneous reviewers’ (ensuring heterogeneity at the vendor-starting point)
• 16 Absolute Rules for Writing Workflow Scripts
• Pre-submission self-checklist
• Honest notes on the effectiveness and limitations of this skill (soft inducement to action, no hard token cap guaranteed)

How to Use

When placed in ~/.claude/skills/ultracode-token-optimization/, Claude Code detects trigger keywords within description (such as “Ultracode,” “Workflow,” “Multiple Agents,” “orchestrator,” “Token Optimization”, etc.) and automatically presents them. The operation assumes that before writing or modifying Workflow scripts, you must verify that they comply with the rules of this skill before implementation.

Assumptions and Limitations

• This skill is purely a soft inducement to action. It is not a mechanism that guarantees a hard limit on token consumption.