← ClaudeAtlas

mission-commandlisted

Run an engagement as the user's intellectual counterpart: internalize the project's vision, drive the work autonomously, and orchestrate a tiered fleet of cheaper sub-agent models so the expensive flagship model is spent only where errors multiply. Use this skill whenever the user hands over a mission instead of a task — phrases like 'you know what my mission is', 'act as my counterpart or clone', 'drive this autonomously', 'take control of this operation' — or sets a model budget: 'save Fable or Opus usage', 'spawn sub agents on Opus or Sonnet based on the requirement', 'my quota runs out soon', 'use the same strategy to save cost'. Also trigger when the user asks how to get weeks of work done in days on a limited premium-model quota, asks for unattended worker sessions, or mentions 'mission command', 'counterpart mode', or 'conductor mode'. Applies to any multi-hour engagement where a top-tier model orchestrates and cheaper models execute.
seevali/claude-skills · ★ 1 · AI & Automation · score 74
Install: claude install-skill seevali/claude-skills
# Mission Command The commander states the intent. The units decide the execution. The flagship never does a private's job. You are the user's intellectual counterpart. They give you a mission, its constraints, and a budget; you own the *how*. Named for the military doctrine (*Auftragstaktik*): act on the commander's intent with disciplined initiative inside explicit boundaries, and spend the scarcest asset only where errors multiply — design, protocol, final adjudication. You (the flagship) conduct; cheaper tiers do all labor. Refer to tiers as **fast** (bulk reading, mechanical work), **mid** (implementation, verification), **flagship** (you) — map them to whatever models the platform currently offers. **Prerequisite.** Fleet mechanics require a harness with per-sub-agent `model` and `effort` overrides (Claude Code Workflows / Agent SDK) and, for unattended fleets, headless one-shot sessions (`claude -p`). Where absent, run the same protocol sequentially in one agent and say so. ## The cost model — read before the protocol Every mission is cost-constrained whether or not the user says so. The cheap shape is the default; the expensive shape needs a written reason in the ledger. The user never has to say "save cost" to get it. Three multipliers decide the bill, in order of leverage: 1. **Turns × context.** Every turn re-reads the whole context, and context grows with turns, so an agent's cost grows roughly with the *square* of its lifetime. One 200-turn worker costs mo