Planner generates PRD from goal

Context and Problem Statement

The agent loop requires a hand-crafted prd.json before it can start. Writing PRDs is the bottleneck — each one requires understanding the codebase, scoping stories to be small enough for a single tool invocation, writing testable acceptance criteria, and formatting as JSON. This takes 10-30 minutes of human effort per loop.

The planner function (ADR-0005) currently does no thinking. It reads the PRD from Redis, finds the next unpassed story, and dispatches. It completes in under a second. Meanwhile, the human is doing the hard cognitive work of decomposition and scoping — exactly the kind of work an LLM is good at.

Decision

Add an optional goal field to the agent/loop.start event. When goal is present and prdPath is absent, the planner spawns an LLM to generate the PRD before dispatching.

Event schema change

"agent/loop.start": {
  data: {
    loopId: string;
    project: string;
    // Existing: read PRD from file
    prdPath?: string;
    // New: generate PRD from goal
    goal?: string;
    context?: string[];  // file paths to read as input (ADRs, docs, issues)
    maxStories?: number; // cap generated stories (default: 6)
    // ... existing fields unchanged
  }
}

Planner flow

goal provided?
  ├─ yes → step: "generate-prd"
  │         - read CLAUDE.md, project structure, context files
  │         - spawn claude with goal + context
  │         - parse output as PRD JSON
  │         - seed to Redis
  │         - continue to story dispatch
  │
  └─ no → existing flow (read prdPath, seed to Redis)

CLI

# New: goal-driven
igs loop start -p ~/Code/project \
  --goal "add Redis client module with typed helpers" \
  --context ~/Vault/docs/decisions/0011-redis-backed-loop-state.md
 
# Existing: file-driven (still works)
igs loop start -p ~/Code/project --prd prd.json

PRD generation prompt

The planner prompt instructs the LLM to:

• Read the goal and context files
• Examine the project structure (package.json, src/ layout, existing tests)
• Generate 3-6 small stories with IDs, titles, descriptions, acceptance criteria
• Each story should be completable by a single tool invocation (codex/claude)
• Acceptance criteria must be verifiable by typecheck + tests
• Output valid JSON matching the existing PRD schema

Guardrails

• maxStories caps generation (default 6, hard max 10)
• Generated PRD is logged to progress.txt for human review
• If generation fails, the loop stops with a clear error — no silent fallback
• Human can still provide prdPath to bypass generation entirely

Consequences

Positive

• Loop startup goes from 10-30 min human effort to a one-liner
• Goals can reference ADRs, issues, or docs as context — the LLM does the decomposition
• Planner becomes the “thinking” step it was always named for

Negative

• LLM-generated stories may be poorly scoped, causing more retries
• Adds ~30-60s to loop start for the generation step
• Generated acceptance criteria may not be as precise as hand-crafted ones

Mitigations

• Start with claude (best at structured output) for generation
• Include examples of good PRDs in the prompt (few-shot)
• Keep prdPath as an escape hatch for cases where human-crafted PRDs are better

References

• ADR-0005 — planner role definition
• ADR-0011 — Redis PRD storage (generated PRDs seed the same way)