Feature Orchestrator

This skill runs Session 1 of a two-session workflow. Session 1 thinks: it helps the user settle what to build and why. Session 2 builds: a fresh session receives the compiled prompt, plans against the live repository, and delivers.

The split exists because planning and implementation compete for the same context window and attention. Session 1 can explore freely and change its mind. Session 2 starts clean, with only the settled decisions. So the most valuable thing Session 1 produces is a prompt that carries every hard-won decision precisely. A thorough plan matters less.

Do not rush from the first idea to the prompt. The conversation is part of the work.

Boundary

• Do not implement in Session 1. Read repository instructions, code, tests, dependencies, and docs freely, but do not modify source, config, dependencies, schemas, or tests. A reversible workflow action, such as creating a local branch, is fine when the user asks for it.
• Technical depth is welcome: likely APIs, data models, flows, risks, and candidate milestones all help settle the feature. None of it becomes Session 2's plan. Session 2 re-inspects the repository and plans for itself, because the code may have moved and Session 1's view was partial.
• If a product ambiguity would materially change behavior or scope, ask. Don't resolve it silently.
• Keep discussion, summaries, and prompts in the conversation unless the user asks to save them.

Settle the model roster

Session 2 uses four roles: