The Recommended Approach

📋 The Four Steps

Step 1 — Research (15–30 min)

Fresh AI context. Full codebase access. Goal: understand before you change.

Prompt examples:

@workspace Trace the [feature] flow from UI to database.
Map every file, class, and method involved. List
dependencies and potential risks.

@workspace What would break if I changed [X]?
List all callers and downstream effects.

Artefact: research.md — dense summary of architecture, dependencies, risks. (~1–2 pages)

Step 2 — Plan (15–30 min)

Fresh AI context. Input: research.md only (information diet). Goal: ordered steps before touching code.

Prompt example:

Given this research document, create an implementation
plan with ordered steps. Each step must include:
- What to change and where
- A verification command (build, test, or lint)
- What could go wrong

Artefact: plan.md — numbered steps with verification commands + risk register

Step 3 — Implement (30–90 min)

Agent mode or coding agent. Input: plan.md. Harnesses running after every change.

Prompt example:

Follow this plan step by step. After each change:
run the build, run the tests, fix any failures before
moving to the next step. Do not skip steps.

Artefact: Working code, all tests green, handoff.md if handing off to someone else

Step 4 — Review (15–30 min)

Human + AI review. Validate against the original intent, not just the code.

Prompt example:

Review this PR against the original plan.md.
Flag any deviations, missing test coverage,
or architectural concerns.

Artefact: Approved PR with ADC record

✅ PR Checklist

Before merging, verify:

All tests pass (unit + E2E if applicable)
Build succeeds in CI
research.md and plan.md committed to ADC folder
Provenance tagged: AI-generated code marked EXTRACTED vs INFERRED
No secrets, credentials, or magic strings introduced
Copilot added as PR reviewer (comments addressed or acknowledged)
At least one human reviewer approved
Architecture constraints in copilot-instructions.md not violated

Roles & Responsibilities

Role	Responsibility
Developer	Runs the RPIR flow. Owns the research and plan quality.
Tech Lead	Reviews plan.md before implementation starts. Validates architectural decisions.
Reviewer	Reviews PR against plan.md intent, not just code style. Checks ADC record.
Agent (Copilot)	Executes within constraints. Self-corrects from harness feedback. Never approves — only suggests.

Quick Commands

Command	When to Use
`@workspace /explain #file`	Research phase — understand a file's role
`@workspace` + question	Research phase — explore architecture
`/tests #file`	Before implementing — establish baseline coverage
`/fix #selection`	Implement phase — fix errors inline
`/doc #function`	Post-implement — document what you built
Agent mode (Ctrl+I)	Implement phase — multi-file changes

⚡ Key Principle

Each phase gets a fresh context with only the artefacts from the previous phase. Don't carry a long conversation through all four steps — start fresh each time. This is the 'information diet' that prevents hallucination and drift.

The Recommended Approach

🎯 The Process at a Glance

Use when

Skip when

📋 The Four Steps

Step 1 — Research (15–30 min)

Step 2 — Plan (15–30 min)

Step 3 — Implement (30–90 min)

Step 4 — Review (15–30 min)

✅ PR Checklist

Roles & Responsibilities

Quick Commands

Self-Paced Resources