Improve

Read SYSTEM.md, current objectives, source registry, decision journal, relevant captures, and recent healthchecks.

The goal is one useful improvement per run, not a broad transformation plan.

Step 1 — Select one problem

Prefer a problem that is:

• repeated,
• expensive in time or errors,
• tied to a current objective,
• and narrow enough to test.

If the user already supplied a specific problem, use it unless evidence shows a prerequisite is missing.

Step 2 — Define the before and after

Capture:

• current process,
• current pain or failure mode,
• desired output,
• one measurable success signal,
• who reviews or owns the result.

Step 3 — Simplify before automating

Ask:

• Can any step be removed?
• Can inputs be standardized?
• Can a deterministic rule solve part of it?
• Which parts genuinely need language-model reasoning?
• Which decisions should remain human-owned?

Step 4 — Map the workflow

Write:

• trigger,
• inputs and authoritative sources,
• ordered operations,
• decisions/branches,
• output destination,
• error path,
• privacy or permission constraints.

Step 5 — Choose the smallest implementation

Use the least complex option that meets the goal:

1. checklist or template,
2. reusable prompt,
3. deterministic script,
4. AI-assisted workflow with review,
5. multi-step agent only when the earlier options are insufficient.

Step 6 — Define acceptance

Before building, specify:

• sample input,
• expected output,
• what must never happen,
• how a human will verify quality,
• what evidence is required before increasing autonomy.

Deliver

Record the decision in journal/decisions.md.

Then create or update one artifact under workflows/ that implements the chosen approach. Preserve unrelated files.

End with:

• what changed,
• how to test it,
• what evidence would justify the next iteration.