Create PRD

Quick Start

1. Copy the template: assets/prd-template.md

2. Fill it out in the target dossier:

   • Single-PRD dossier: prd.md
   • Multi-PRD dossier:
     • Overall / spine PRD at the dossier root: prd-overall.md
     • Slice PRDs under prds/: prds/<slice_id>_<slug>/prd.md

   Notes:

   • Lowest-level PRD filenames are always prd.md (no prd-slice-* filenames).
   • Use a short kebab-case slug like auth-token-rotation or invoice-pdf-export.

3. Pull all details from shaping inputs into the PRD so it’s implementable without extra context (no information loss)

PRD Slicing Rubric (Reusable)

Rules of thumb (thin PRDs)

• One PRD should deliver one user-visible affordance (or one backend capability with a measurable, observable effect), not a whole workflow.
• Each PRD must have measurable acceptance criteria tied to at least one replayable verification pack (fixture data, synthetic dataset, scripted scenario, etc.).
• Keep PRDs vertical-ish: UI + a thin API/data change where needed, but avoid building platform layers unless they directly unblock the next slice.
• If a PRD introduces a new failure mode, it must also introduce the UX to surface it (no silent failures).
• Prefer scaffold then replace: ship a deterministic baseline behind stable contracts first, then swap in ML/OCR/retrieval/LLM (or other complex components) behind the same interface.

Examples (good slices)

• "UI affordance X triggers state Y, persists, and is visible after refresh."
• "API returns payload + stable checksum/hash so clients can detect mismatches."
• "Verifier rejects records on checksum mismatch and surfaces a user-visible error state."
• "Parser extracts a single structured list into a table with row-level provenance."

Examples (bad slices)

• "Build ingestion + parsing + retrieval + drafting + verification + export" (too many risks bundled).
• "Implement an autonomous multi-agent copilot" (non-deterministic + typically too broad for a single slice).
• "Add full template/customisation system" (platform work without proving the core loop).

Red flags a PRD is too fat

• Touches 3+ major subsystems/components in one go.
• Needs more than 1-2 new schemas or widespread data model churn.
• Acceptance criteria can't be tested against a pack/fixture/script.
• Contains multiple hard unknowns at once (new parsing + matching + verification together).

Minimum PRD structure (every PRD)

The minimum structure is defined by assets/prd-template.md and must include:

• Introduction/Overview (problem + goal)
• Goals
• User stories
• Functional requirements (numbered)
• Non-goals (out of scope)
• Success metrics
• Open questions

Additionally, every PRD must include:

• Acceptance criteria + verification plan (pack/fixture/script)
• Failure states + user-visible UX (no silent failures)
• Metrics/logging (at least 1 signal)
• Rollback/disable path (feature flag or safe default)
• A lightweight "Sources" section pointing to shaping inputs (brief/breadboard/spike/risk-register)

Inputs

One or more of the following:

• Problem statement + target users
• Success metric
• Constraints, risks, dependencies
• Any existing verification packs/fixtures/scripts (or willingness to create a minimal one)

No Context Loss (Requirement)

If any of the following shaping files exist, no information should be lost from them:

• brief.md
• breadboard-pack.md
• spike-investigation.md
• risk-register.md

Guidance:

• Prefer integrating shaping details directly into the PRD sections (requirements, risks, constraints, UX, verification).
• Add a short "Sources" section with links/paths to the shaping files you used.
• If a detail does not fit cleanly into a PRD section, add it to the appendix as notes or key excerpts. Avoid dumping large verbatim blocks by default.
• Each PRD must be implementable without needing additional context outside the PRD.

Try find this information:

• brief.md
• breadboard-pack.md
• spike-investigation.md
• risk-register.md

Where unclear, use the ask-questions-if-underspecified skill.

Outputs

• Single-PRD dossier: prd.md in the dossier root.
• Multi-PRD dossier:
  • Overall / spine PRD in the dossier root as prd-overall.md.
  • Slice PRDs under prds/<slice_id>_<slug>/prd.md.

Steps

1. Ask 3-5 clarifying questions (answerable as 1A, 2B) that force a thin slice:

   • What is the single user-visible affordance (or single observable backend capability)?
   • What is the primary observable effect (what changes, where can we see it)?
   • What pack/fixture/script will we use to verify acceptance criteria?
   • What new failure modes does this introduce, and how should UX surface them?
   • What is the rollback/disable path (flag/default-off/fallback)?

2. Run the slicing rubric:

   • If the scope triggers a red flag, propose a split into PRD-1/PRD-2/... and proceed with the smallest slice that unblocks the next.

3. Use the PRD template at assets/prd-template.md.

4. Fill sections (short, explicit; keep the template headings):

   • Introduction/Overview + goals
   • In scope + non-goals (out of scope)
   • User stories (US-001...), one sprint-sized
   • Functional requirements (FR-001...)
   • Acceptance criteria + verification plan per story (measurable; tied to a pack/fixture/script)
   • Failure states and UX (no silent failures)
   • Metrics/logging + rollback/disable path
   • Risks + dependencies
   • Success metrics + open questions
   • Sources: link to shaping inputs used
   • Appendix (optional): shaping notes / key excerpts when needed (see "No Context Loss")

5. Ensure criteria map to verification; split oversized stories. If generating multiple PRDs:

   • Create the overall/spine PRD at the dossier root as prd-overall.md.
   • Create each slice PRD under prds/<slice_id>_<slug>/prd.md (one folder per slice).

   Each PRD must be independently implementable and must not lose shaping input information.

Verification

• PRD delivers one affordance / one observable capability.
• Acceptance criteria measurable, testable, and tied to at least one pack/fixture/script.
• Acceptance criteria mapped to verification (tests/build/smoke/scripts).
• Failure modes surfaced in UX (no silent failures).
• At least 1 metric/log signal is specified.
• Rollback/disable path is explicit (flag or safe default).
• Stories small, priority ordered.
• Location matches docs/AGENTS.md.
• If there are multiple PRDs for a dossier:
  • the overall/spine PRD is prd-overall.md in the dossier root
  • each slice lives under prds/<slice_id>_<slug>/prd.md (folder-per-slice; lowest-level PRD filename is always prd.md)
• Sources section links to shaping inputs used.
• No information is lost from shaping inputs (details are integrated into the PRD or captured in the appendix).