Four Part Prompt

<!-- VCB:BEGIN_TOPIC id=vcb.chapter.four_part_prompt version=0.1.0 -->
---
id: vcb.chapter.four_part_prompt
title: 'Chapter 7 — The Four-Part Prompt: Goal, Context, Constraints, Done When'
type: chapter
chapter_number: 7
version: 0.1.0
last_verified: '2026-06-08'
next_review_due: '2026-12-08'
review_cadence: semiannual

audiences:
- human
- ai_coach

applies_to:
- Codex App
- Codex CLI
- Codex IDE Extension
- Codex Cloud
- GitHub
- AI coach workflows

source_status:
  official_openai: true
  official_vendor: false
  community_field_practice: false
  speculative: false

evidence_level: E0_OFFICIAL_DOCS
evidence_basis: VCB_SYNTHESIS_STABLE_ENGINEERING_PRACTICE
source_kind: official_docs_plus_maintainer_synthesis

stability:
  principle: AGENTIC_PRINCIPLE
  surface: MODERATE
  pricing: STABLE

budget_profiles:
- plus_constrained
- pro_high_throughput
- api_pay_as_you_go
- team_shared_budget

cost_postures:
- cheapest_reliable
- balanced
- fastest_possible
- unattended_high_throughput
- production_quality
- disposable_prototype

project_phases:
- prototype
- mvp
- production_build
- maintenance
- major_refactor
- emergency_hotfix

attention_modes:
- continuous_supervision
- periodic_check_in
- low_attention_review_later
- unattended_requires_isolation

shortcut_profiles:
- vcb.shortcut.vague_prompt
- vcb.shortcut.one_big_prompt

durable_principles:
- A useful Codex prompt names the outcome, relevant context, constraints, and observable completion criteria.
- Prompt structure reduces guessing and rework.
- Prompt detail should scale with ambiguity and risk.
- Prompt wording is not proof; the diff and verification evidence still matter.

likely_to_change:
- Codex Goal mode mechanics
- Codex Plan mode mechanics
- surface-specific context capture
- slash-command names and UI controls
- model-specific prompt sensitivity

obsolete_when:
- Codex reliably infers goal, context, constraints, and done-when for production tasks without explicit user guidance.
- Official OpenAI guidance replaces the goal/context/constraints/done-when pattern with a materially different default task contract.

related_topics:
- vcb.chapter.plan_first_code_second
- vcb.chapter.context_management
- vcb.chapter.acceptance_criteria
- vcb.chapter.codex_mental_model
- vcb.chapter.git_discipline
- vcb.prompting.four_part_prompt
- vcb.shortcut.one_big_prompt
---

> Summary:
> The four-part prompt is the default work order for Codex: goal, context, constraints, and done when. It turns a vibe into an executable, reviewable task.

## Quick Navigation
- For the Human
- For the AI Coach
- Shortcut Reality
- Budget and Constraint Notes
- Stable Core
- Volatile Surface
- Obsolescence Watch
- Evidence and Sources
- Related Topics

<!-- VCB:BEGIN_SECTION id=vcb.chapter.four_part_prompt.human -->
## 1. For the Human: Plain-Language Concept

### What this means

A Codex prompt is not a wish. It is a work order.

A strong work order has four parts:

| Part | Plain meaning | Example |
|---|---|---|
| Goal | What should exist after the task? | “Add a password reset flow.” |
| Context | What should Codex inspect or know first? | “Use the existing auth routes in `app/auth/*` and this error log.” |
| Constraints | What must Codex preserve or avoid? | “No new dependencies. Keep the API response shape.” |
| Done when | What proves the task is complete? | “Invalid tokens fail safely and auth tests pass.” |

Think of it like ordering food. “Make me something good” can work at a food truck. It is not enough when you need allergy-safe catering for a hundred people. Codex can improvise, but higher-risk code should not depend on improvisation.

### Why it matters

Vague prompts create invisible assumptions. Codex fills gaps using repo patterns, common framework habits, and guesses. Sometimes the guesses are useful. Sometimes they quietly change behavior you cared about but never stated.

A four-part prompt narrows the target: