| name | improve |
| description | Run the full review-first improvement workflow for a target file, directory, or glob by dispatching the five independent improve components, consolidating findings, then applying only user-selected changes. |
Instructions
This is the orchestrator. For a single lens, call the component skills directly:
$improve-analysis
$improve-upstream
$improve-simplification
$improve-security
$improve-prior-art
Read references/target-and-agent-contract.md before dispatching any component. Read references/consolidate-and-act.md before presenting findings or editing files.
Workflow
- Parse
$ARGUMENTS with the shared argument contract.
- Resolve the target to the exact file list with the shared target-resolution rules.
- Read the target lightly and produce the shared context brief:
- resolved target paths
- what the target appears to do
- 2-3 concrete usage scenarios
- test-file flag, when applicable
- optional
<user_note> emphasis, when applicable
- Inspect the tokens after the target:
- If every trailing token exactly matches one of
analysis, upstream, simplification, security, or prior-art case-insensitively, dispatch only those components.
- Otherwise, dispatch all five components and pass the whole trailing-token remainder as
<user_note> emphasis.
- Dispatch the selected components in parallel as read-only subagents. Tell each subagent to follow the matching component skill:
analysis -> .agents/skills/improve-analysis/SKILL.md
upstream -> .agents/skills/improve-upstream/SKILL.md
simplification -> .agents/skills/improve-simplification/SKILL.md
security -> .agents/skills/improve-security/SKILL.md
prior-art -> .agents/skills/improve-prior-art/SKILL.md
- Consolidate reports with
references/consolidate-and-act.md.
- Stop after presenting findings. Do not edit until the user chooses which numbered findings to apply.
If a component returns empty, off-topic, errors out, or fails to return, note it in the consolidation and proceed. Do not re-dispatch unless the user asks.