Hai PRD
For Chinese readers, see SKILL.zh_CN.md. The English SKILL.md is the execution source of truth.
Overview
Handle all PRD-related work: deciding whether a PRD is even needed, choosing PRD granularity, splitting or merging PRDs, drafting a PRD from a product problem, and refining an existing one.
The non-obvious framing: a PRD is not the default artifact. Not every idea, field, API, configuration change, or technical decision earns one. Pick the mode that fits the request and do not wrap work in a PRD when there is no product boundary or acceptance to prove.
Core Principle
A PRD is a product argument with verifiable acceptance — not a feature list. Three rules carry the rest:
- Acceptance must be provable. One PRD should resolve one clear outcome, one main user journey or product boundary, and acceptance criteria that prove it shipped.
- Boundary follows the user's mental model, not the tech. Granularity tracks user journeys and independent acceptance — not technical layers, teams, or files.
- Fix the boundary before the wording. If the PRD boundary is wrong, split, merge, or reframe it before polishing local sentences; edit an existing PRD as a system where goal, scope, behavior, acceptance, risks, and rollout must agree.
Modes
1. Need PRD
First decide the artifact type — a PRD is one option among several:
| Situation | Better artifact |
|---|---|
| User or business behavior changes and goal, scope, and acceptance need alignment | PRD |
| Execution goal, phases, and verification are the main need | hai-goal |
| Architecture, module boundaries, or technical tradeoffs are the main need | hai-architecture |
| The main question is whether the idea is worth doing at all | hai-idea |
| The work is a code task, bugfix, or test-first implementation | hai-tdd or task note |
| The work is documentation consistency auditing | hai-audit-docs |
Ask:
- Is there a user-, buyer-, operator-, maintainer-, or business-visible behavior change?
- Does the work need explicit in-scope and out-of-scope boundaries?
- Does it need acceptance criteria to prove it was done correctly?
- Are there product-level options, risks, constraints, or rollout concerns to record?
- Would the team disagree on goal, scope, or acceptance without a PRD?
If most answers are no, do not force a PRD. When auditing a set of PRDs, also check for duplicate boundaries, circular dependencies, or broken user journeys across them.
2. Design PRD
Turn a feature idea or product problem into PRD-level requirements:
- Gather context. Target user, product surface, current behavior, desired outcome, constraints, assumptions; plus existing PRDs, docs, screenshots, user flows, issues, data, or adjacent specs.
- Frame the requirement. Write the problem in product language; define the target outcome and non-goals; identify the user journey or workflow boundary; separate facts, assumptions, and open questions.
- Shape scope. Decide what belongs in this PRD versus another PRD, design doc, goal, or task. Do not bundle unrelated cleanup, platform work, or refactors unless they directly affect product behavior or acceptance.
- Define behavior and acceptance. Describe expected behavior by scenario; add observable acceptance criteria; include edge cases, failure states, permissions, empty states, data dependencies, and rollout only when relevant.
3. Craft PRD
Refine an existing PRD as a whole product argument:
- Read the full document first. Identify title, target user, goal, journey, scope, non-goals, acceptance, risks, and open questions; mark contradictions, duplication, stale sections, local patches, and unreasonable content.
- Diagnose document drift using this taxonomy:
- Goal drift: sections optimize for different outcomes.
- Scope drift: requirements pull in unrelated work.
- Conflict drift: behavior, state, permissions, data, or rollout disagree across sections.
- Acceptance drift: acceptance criteria cannot prove the target outcome.
- Solution drift: implementation choices replace product requirements.
- Local-edit drift: one section improves while the whole PRD becomes worse.
- Reframe before rewriting. State the corrected target outcome and the PRD boundary; decide whether to preserve, reorganize, split, merge, or rewrite.
- Craft the document. Edit directly when the user provides a file; preserve useful content but move it to the right section; remove unsupported, duplicated, obsolete, or out-of-scope content; explain why removal beats rewriting or moving when content is deleted.
4. Scope PRD
Judge PRD granularity by user mental model and independent acceptance, not technical modules or teams. Run the six tests — this is the decision core:
| Test | Question | Signal |
|---|---|---|
| Press Release | Can one sentence express one clear user value? | Multiple unrelated values lean split |
| Independent Value | Does this PRD deliver value on its own? | Independent value can split; no value leans merge |
| Independent Acceptance | Can it be accepted independently? | Independent acceptance can split |
| Domain Language | Does it share the same core entity, terminology, and lifecycle? | Same entity leans merge; different contexts lean split |
| User Journey | Does splitting break the user journey? | Broken journey means do not split |
| Time Appetite | Is the scope controlled? | Too large should split or cut; too small should merge |
When the tests point toward split or merge, read references/scope-tiers.md to pick a concrete
cut: the preferred split dimensions, the splits to avoid (technical layer / CRUD / team ownership,
because no such fragment is independently acceptable), the three-tier classification (entity
lifecycle / aggregate view / platform extension), and the new-feature decision flow.
Workflow
- Pick the mode from the request: necessity check → Need PRD; draft from a problem → Design PRD; refine an existing doc → Craft PRD; split/merge/granularity → Scope PRD. A request may span modes (e.g. necessity then scope).
- Run that mode's procedure above. Apply the Core Principle throughout — fix the boundary before the wording.
- For split/merge questions, run the six tests, then consult
references/scope-tiers.mdfor the concrete cut and classification. - Produce the output for the active mode (see below). When a worked run helps, read
references/worked-example.mdfor an end-to-end split decision and an end-to-end craft/diagnosis.
Output
Read references/output-template.md and fill only the active mode's sections. Do not mechanically
emit every mode or duplicate the template here. A necessity check can be brief; a draft or repair
must carry enough product behavior and acceptance detail to stand alone.
Use a different skill when
Routing away from a PRD is the Need-PRD decision itself — see the table in Mode 1. In short: settle execution phases/todos in hai-goal, module boundaries and tradeoffs in hai-architecture, whether the idea is worth doing in hai-idea, a code/bugfix/test-first task in hai-tdd, what fields an entity should have and where they live in entity-model-auditor, documentation consistency in the doc-audit skills, and an anchor-first rewrite of a drifted non-PRD document in hai-rewrite-doc. Return to PRD work only when a product boundary and provable acceptance actually need to exist.
What this skill is NOT
- Not project management — do not assign owners or dates unless asked.
- Not a wrapper that turns every idea into a PRD — if there is no product boundary or acceptance need, recommend the right artifact from the Need-PRD table instead of forcing a PRD.
