Small Model Discipline

Small-Model Discipline

When to engage

• Task touches >1 file or >1 concept
• User reports wrong/incomplete prior output
• Using a fast or smaller model
• Requirements are ambiguous

Mandatory loop

1. RESTATE goal in one sentence
2. PLAN max 5 bullets (files + actions)
3. VERIFY plan against repo (grep/read) — 1 pass
4. IMPLEMENT one bullet at a time
5. CHECK diff matches goal only
6. STOP — don't add extras

Do not skip step 3. Do not implement all bullets in one blind pass.

One thing per turn (internal)

For complex work:

• Turn A: explore + plan (no edits)
• Turn B: implement core change
• Turn C: fix lints/tests if needed

If the environment allows single-turn completion, still execute the loop mentally before editing.

Verification hooks

Before marking done:

• Grep shows no leftover TODO/debug prints you added
• Imports resolve; no orphan new files
• Behavior matches restated goal — not a superset
• Diff size plausible for task (see minimal-diff)

Error recovery

On failure:

1. Read the error message literally — don't guess.
2. Change one variable (one file or one hypothesis).
3. Retry once. If still failing, report blocker with evidence.

No shotgun edits across unrelated files.

Pair with

All other skills in this pack — especially token-efficient-exploration and minimal-diff.

Escalation

If after 2 focused attempts the task needs architectural decisions, ask one clear question instead of generating large speculative code.