cut-it

Cut It

Take a plan that already exists and cut it into clean slices — ordered phases an AI agent can pick up and execute one at a time. Each slice is a workable, verifiable chunk that leaves the codebase running. The flavor is orc; the slices stay clean and precise — an agent must execute them cold.

1. Find the plan — do not invent one

Work from a plan that already exists. Look in this order:

1. The plan in this conversation — an approved design, plan-mode output, a decision just reached.
2. A file or ticket the user points to — PLAN.md, a GitHub issue, a PRD, a doc.
3. If none is clear, ask the warchief to point at it or paste it. Never fabricate the plan.

Read the whole thing first. Name the end goal, the moving parts, and what depends on what.

2. Cut into slices

Rules for every cut:

• Vertical, not horizontal. Each slice delivers something that works end-to-end, not a half-built layer.
• Dependency-ordered. Foundations first. No slice depends on a later slice.
• One agent run each. Small enough to execute and verify in a single focused session; big enough to matter. Split a fat slice; merge a trivial one.
• Always green. Each slice leaves the codebase building, tests passing, app running — a safe commit point.
• Self-contained. Readable cold, with real file paths and commands. No reliance on this conversation.

3. Write each slice

## Slice N — <short imperative title>
**Goal:** one line — what this slice delivers.
**Depends on:** Slice X, Slice Y   (or: none)
**Touches:** files / modules / areas to create or change.
**Steps:**
  1. concrete action
  2. concrete action
**Done when:** the explicit, verifiable check — a test passes, a command prints X, the behavior works.
**Out of scope:** what NOT to do here, so the agent stays in its lane.

Done when is the most important line — it is how the agent knows the slice is finished.

4. Hand it off

• Open with the battle order: numbered slice titles and their dependencies, so the shape is clear at a glance.
• Then the full slices, in execution order.
• Offer to save them to a file (PLAN.md or .claude/plans/<name>.md) so an agent can pick them up, or to feed them to /to-issues.

Guardrails

• Slice the real plan. If it is vague, sharpen it or ask — do not paper over gaps.
• Every slice independently verifiable, with no forward dependencies.
• Concrete over clever: real paths, real commands, real acceptance checks.
• If the plan is too big for clean slicing, say so and propose where to split the work.