Prompt Pipeline

# Prompt Pipeline Design

The prompt pipeline is the engine of SDD. Each numbered prompt drives one phase of the Hunt lifecycle (Spec, Plan, Implement, Iterate, Monitor). Prompts are structured markdown files that instruct an AI agent to perform a specific phase, with detailed information delegated to specs, plans, and reference materials.

**Core principle:** Prompts should be as lightweight and systemic as possible. They define the *process*, not the *content* -- specs and plans hold the content.

---

## 1. Greenfield Pattern (3-Prompt Pipeline)

For new projects starting from reference materials (PRDs, language specs, design docs, research).

```
Pipeline Flow:
  refs/ ──> [001] ──> specs/ ──> [002] ──> plans/ ──> [003] ──> src/ + tests/
                                   ^                     |
                                   |                     |
                                   +── impl/ <───────────+
                                   (bidirectional flow)
```

| Prompt File | Lifecycle Stage | Reads From | Writes To | Description |
|-------------|----------------|------------|-----------|-------------|
| `001-generate-specs-from-refs.md` | **Spec** | `context/refs/` | `context/kits/` | Reads reference materials, decomposes into domain-specific specs with cross-references and testable acceptance criteria |
| `002-generate-plans-from-specs.md` | **Plan** | `context/kits/` + `context/impl/` | `context/plans/` | Reads specs plus implementation progress, creates framework-specific plans with feature dependencies, test strategies, and acceptance criteria |
| `003-generate-impl-from-plans.md` | **Implement** | `context/plans/` + `context/kits/` | `src/`, `tests/`, `context/impl/` | Implements the highest-priority unblocked work from plans, runs tests, updates implementation tracking |

### Key behaviors

- **Prompt 001** runs once or a few times to stabilize specs. It reads `context/refs/` and produces `context/kits/`.
- **Prompt 002** reads specs and any existing implementation tracking (`context/impl/`). It produces plans that sequence the work.
- **Prompt 003** reads plans and specs, implements code, runs validation gates, and updates `context/impl/` with progress.
- **Prompts 002 and 003 modify each other's files.** This bidirectional flow is expected and healthy -- it is how the system self-corrects.

### Example prompt 001 structure

```markdown
# 001: Generate Specs from Reference Materials

## Runtime Inputs
- Framework: {FRAMEWORK}
- Build command: {BUILD_COMMAND}
- Test command: {TEST_COMMAND}

## Context
Read all files in `context/refs/`. These are the source of truth.

## Task
Decompose the reference materials into domain-specific specifications:
1. Create `context/kits/cavekit-overview.md` as the index file
2. Create one `context/kits/spec-{domain}.md` per domain
3. Each spec must include: Scope, Requirements with Acceptance Criteria, Dependencies, Out of Scope, Cross-References

## Exit Criteria
- [ ] All domains from reference materials have corresponding spec files
- [ ] Every requirement has at least one testable acceptance criterion
- [ ] cavekit-overview.md indexes all spec files with one-line summaries
- [ ] Cross-references link related specs

## Completion Signal
<all-tasks-complete>
```

---

## 2. Rewrite Pattern (6-9 Prompt Pipeline)

For projects that start from existing code that must be reverse-engineered into specs before building a new implementation.

```
Pipeline Flow:
  old-code ──> [001] ──> reference/ ──> [002] ──> specs/ ──> [003] ──> validated specs