work

The compound engineering loop — five discrete stages that systematically reduce rework by front-loading clarity and back-loading lessons. Each stage can be run independently or in sequence.

Git Root Detection

Before running any command, detect the git root:

git rev-parse --show-toplevel

Commands

Command	Description
brainstorm	Clarify scope through dialogue, surface key decisions and open questions
plan	Research codebase, write concrete implementation plan
work	Execute the plan, track progress in task backend
review	Evaluate output, identify lessons, decide accept/rework
compound	Document lessons, update rules/templates/docs
spec <topic>	Create outcome spec (overview, decisions, steps, edge cases)
audit	Review current branch for blindspots; fix now / needs spec / create issues

$ARGUMENTS Routing

$ARGUMENTS	Action
help	→ commands/help.md
brainstorm	→ commands/brainstorm.md
plan	→ commands/plan.md
work	→ commands/work.md
review	→ commands/review.md
compound	→ commands/compound.md
spec	→ commands/spec.md
audit	→ commands/audit.md
(empty)	→ commands/help.md

Stage Task Naming

Each stage creates or updates a task in the configured backend (if tasks != none). Tasks follow the naming convention (no stage numbers):

YYYY-MM-DD, <Stage>

e.g. 2026-02-28, Brainstorm, 2026-02-28, Plan, 2026-02-28, Compound — <topic>

Tasks are created under the relevant workstream epic if one exists.

Frustration Detection

If the user shows signs of repeated fixes, confusion, or rework — proactively suggest running work compound even mid-stream. The value of compound is in capturing lessons before they fade.

Design Note

80% of the compound loop's value comes from plan and review. work is execution. Do not skip plan — scope creep and rework almost always trace back to an unclear plan.