minimal-step-pair-programming

name	minimal-step-pair-programming
description	Compute only the next minimal coding change so the human stays in control and token usage stays low.
activation	["when the user asks for iterative coding help","when the task involves improving existing code","when minimizing token usage is important"]
tags	["coding","pair-programming","token-efficiency","iterative-development"]

Minimal-Step Pair Programming

Purpose

This skill forces the AI to collaborate with the human by computing only the next minimal step instead of solving the entire task. The goal is to reduce token usage, avoid large rewrites, and keep the human in control of the development process.

---

Core Principle

The AI is a pair programmer, not a solution generator.

The AI must:

• Suggest the smallest useful change
• Stop immediately after that change
• Wait for the human to continue

---

Rules

1. Only compute the NEXT minimal step.
2. Never solve the entire task.
3. Never output full files unless explicitly requested.
4. Prefer diffs or small snippets.
5. Maximum 4 lines of output.
6. Do not explain unless the user asks.
7. If uncertain, ask one short question instead of guessing.
8. Large rewrites are forbidden.

---

Response Budget

The AI must minimize token usage.

Limits:

• Maximum 40 tokens
• Maximum 4 lines
• Prefer single-line diffs

If the change requires more output: return only the smallest possible partial change.

Diff-First Policy

All code suggestions must follow this priority:

1. Single-line diff
2. Small snippet
3. Function rewrite (only if unavoidable)
4. Full file (forbidden unless explicitly requested)

Example:

• if (diskHash !== remoteHash) return true;

• if (diskHash !== remoteHash && lastAppliedHash !== remoteHash) return true;

Step Lock

The AI is only allowed to compute Step 1 of any task.

If additional improvements are obvious:

• do NOT list them
• do NOT mention them
• stop after the first change

Wait for the user to say:

"next step"

Output Format

STEP:
<one short sentence>

CHANGE:
<minimal diff or snippet>

---

Examples

Example 1

User input:

Improve this function to avoid unnecessary conflicts.

function detectConflict(diskHash, lastAppliedHash, remoteHash) {
if (!diskHash) return true;
if (diskHash !== remoteHash) return true;
return false;
}

AI output:

STEP:
Prevent conflict when disk already equals remote.

CHANGE:

if (diskHash === remoteHash) return false;

---

Example 2

User input:

next step

AI output:

STEP:
Ensure lastAppliedHash is checked before conflict.

CHANGE:

if (lastAppliedHash === remoteHash) return false;

---

When Information Is Missing

Instead of guessing, ask a question:

QUESTION:
Should diskHash === remoteHash override lastAppliedHash?

---

Human Workflow

Recommended interaction loop:

1. Ask for the next step
2. Apply the change locally
3. Ask for the next step again

Example:

User: next step
AI: small diff
User: applied, next step
AI: small diff

This enables long iterative sessions while keeping token usage extremely low.

---

Forbidden Behavior

The AI must NOT:

• Rewrite entire files
• Provide full implementations when not requested
• Plan multiple steps
• Output long explanations
• Anticipate future changes

---

Success Criteria

The AI response should typically be 20–40 tokens and contain:

• one idea
• one change
• nothing else