PRD Implementation Ready

What This Skill Does

Use this skill to figure out the product with the user before writing the document. The output must be readable for humans and complete enough that a developer or AI can implement it without asking another business question.

Use it for PRDs, BRDs, functional specs, SaaS/admin workflows, C-end/content products, tool SaaS, consumer apps, finance/payment/settlement flows, product iteration from v1 to v2, or any requirement where state, data, permission, page UX, notification, non-functional product behavior, or cross-module linkage can break implementation.

When NOT to use it: one field label, copy edit, visual-only tweak, pure technical implementation task, or a single-step requirement with no state/data/source/cross-role risk.

Core bar:

• Context-first business understanding: user-provided context is the first source of truth. Build a context digest before applying the workflow.
• Product quality plus completeness: a complete bad product is still bad. Research the domain, test whether the flow feels natural, and improve clumsy UX instead of only filling tables.
• Reader-first module PRD structure: write Module purpose, Main flow, Page presentation, Key operations, Risks and exceptions, Page implementation details, Server-only automatic flows, and acceptance in that order.
• Close pending decisions when safe: use business anchors, source data, existing behavior, or low-risk best-practice defaults. Leave pending only when the user explicitly says it is undecided, sources conflict, or no safe default exists.
• Do not invent terms: use the user's wording, existing system labels, or confirmed business terms.
• Human-readable wording: PRD正文 must stay business-readable and aligned with the requirement. Do not replace user terms with abstract, decorative, or void-created terms.
• Page/document term alignment: page labels, PRD terms, business objects, calculation fields, and fund/resource destinations must use one mapped vocabulary before final wording.

Default first response shape: context digest based on the user's context; deep business understanding plan; one-sentence business goal; known/unknown scope, users, objects, and risks; initial flow shape; likely small flows from observation lenses; inferred high-risk control loops for repeated/cyclic operations; 3-6 confirmation questions with recommended defaults.

How To Run It

references/stage-gates.md is the authoritative workflow. In plain text: stage-gates.md is the authoritative workflow. The progressive workflow summary is:

1. If the user asks how to use the skill, gives a mixed request, or the next action is unclear, choose the usage path from references/usage-path.md before loading heavy references.
2. Business understanding first: context digest, deep business understanding package, external research, flow awkwardness analysis, micro-flow discovery, exception/reverse understanding, business anchors, and scope.
3. Flow shape next: identify single-line, parallel, master-child, state-machine, event-driven, cycle/periodic, or mixed flow. For parallel flows, name convergence points instead of flattening the work.
4. Navigation and diagrams before detail: the main PRD / 00 file should include a side/module/page/function navigation map and a business flow atlas before detailed rules.
5. Rules after flow: object lifecycle, data flow, source documents, writebacks, idempotency, permissions, exceptions, boundaries, and rollback.
6. Pages after business shape: page task closure, target touchpoint, page carrier, component choice, interaction states, and feedback/recovery.
7. Final writing after confirmation: show confirmed rules, recommended defaults, open high-risk decisions, and out-of-scope boundaries.

Diagram requirements: always consider flow shape, parallel flow convergence, business flow atlas, Mermaid flowcharts, stateDiagram-v2, object status changes, and diagrams before detailed rules when the PRD spans multiple objects/modules or has lifecycle/state risk.

For existing products or v1 to v2 changes, run the Version Evolution Gate before rewriting: old rule, change intent, impact range, old vs new, migration, backward compatibility, release/rollback, and decision record.

For module docs, keep details near the page/action/server flow where they matter. Page implementation details must include a page element inventory, fields, data source, status/writeback, permissions, idempotency, validation, error/empty/disabled states, feedback/recovery, and acceptance. Server-only automatic flows stay inside the module with trigger, input source, processing rule, persisted object, status/writeback, idempotency/retry, failure/compensation, visibility/log, and acceptance.

Reference Router

Read only what the current stage needs.

Gate authority:

• Authoritative gate: references/data-flow.md#post-event-writeback-gate.
• Authoritative gate: references/data-flow.md#resource--value-flow-gate.
• Finance topic index: references/finance-operations.md.
• Unified staged red flags and P0/P1 blockers: references/guardrail-checklist.md.

Guard Rails

Stop and return to the workflow when any of these appears:

• Process-before-context or template-first PRD writing: starting from a generic structure without digesting user-provided context.
• Module writing before the deep business understanding package is complete enough for the selected depth.
• Mainline-only thinking that misses business observation lenses, likely small flows, role handoffs, server-only flows, recovery paths, or page actions.
• Finance/logistics-only examples used for a C-end, consumer app, content/social, or tool SaaS product without the Domain Adaptation Table.
• A v1 to v2 change treated as a fresh PRD without old rule, change intent, impact range, old vs new, migration, compatibility, rollout, rollback, and decision record.
• Thin user stories that omit context, trigger, preconditions, alternate path, success criteria, or emotion/friction.
• Missing product-level non-functional decisions when performance target, SLA, data visibility latency, consistency, retry, degradation, or abuse control changes behavior.
• Scattered notification design: a notification, todo, push, email, SMS, bot message, or webhook appears only as a side effect without trigger event, receiver, channel, dedupe, read/task state, recall/update, preferences, failure handling, and deep link.
• Invented terminology: a new term appears without user wording, existing system labels, confirmed business terms, or a clear recommended naming basis.
• Page/document term drift: the same concept appears under different page labels, PRD terms, formula components, ledger effects, or fund/resource destinations without an explicit mapping and chosen wording.
• Analysis metric used as an action hard gate: attribution, occupancy, or cost attribution blocks settlement/payment/refund/release without reliable source and a named business action consumer.
• Unfiltered source data: a PRD says take data from an external system, other module, upstream object, import, sync result, or configuration without source dataset, include/exclude conditions, source state, scope, period window, availability, unavailable reason, snapshot, and refresh behavior.
• Hidden amount composition source: amount breakdown names cost basis, source total, service fee, adjustment lines, or included detail list without source lineage, drilldown, snapshot, and detail-sum reconciliation.
• Unqualified real-time relationship: amount, list scope, permission, or report dimension says current binding, current relationship, latest configuration, or real-time fetch without relationship source, effective period, anchor time, snapshot, rebuild policy, and correction behavior.
• Field-only page design: columns/buttons are listed but why shown, what decision or action they support, operation surface, disabled reason, and recovery are missing.
• Large-region page design: page presentation stops at page purpose, big areas, or "list + detail" without a complete visible element inventory for page header, summary cards, filters, tabs, table/list, actions, drawers/modals, logs, import/export, and interaction states.
• Created downstream records are listed after an event, but exact persisted field writebacks on existing objects are missing.

For the full staged list, use references/guardrail-checklist.md. Any P0 blocker means stop detailed writing until the issue is closed or explicitly marked pending with a recommended default.