行业知识 · v2.3.0 · 资料核对 2026-10-03

Discovery and PRD Reference

Evidence ladder

Prefer evidence in this order:

Observed user behavior and completed-task data.
Recent concrete examples described by users or frontline staff.
Transaction, workflow, support, and operational records.
Controlled prototype or willingness-to-pay behavior.
Stated preferences and stakeholder opinions.
Team intuition.

Label weak evidence. Do not transform a stakeholder request into a validated user need.

Current-workflow reconstruction

Capture trigger, actor, goal, steps, tools, handoffs, information, time, cost, errors, exceptions, approval, and final outcome. Identify the bottleneck and whether the root cause is information, coordination, policy, capacity, interface, or judgment.

Use recent-event questions:

When did this last happen?
Show or reconstruct each step.
Where did you wait, rework, check, or ask for help?
What made the outcome acceptable?
What happens when the result is wrong?
What information cannot be shared with a system?
Who approves a process change or pays for it?

Opportunity scoring

Score 1–5 and show evidence for:

Problem severity and frequency.
Number/value of affected tasks.
Willingness and authority to change.
Data/context availability and legality.
AI feasibility and measurable output.
Error tolerance and reversibility.
Time to first proof.
Strategic fit and economic upside.
Operational and adoption burden.
Safety, compliance, and reputational risk.

Do not sum blindly. Treat unacceptable risk or unavailable lawful data as gates.

AI-fit tests

Prefer simpler solutions when the task has stable rules, exact outputs, low variance, or cheap deterministic lookup. Consider AI when value depends on unstructured inputs, semantic matching, prediction, generation, or adaptive judgment and when errors can be bounded.

Compare:

Option	Test
Process redesign	Can the bottleneck be removed without automation?
Rules	Can experts express stable conditions and outputs?
Search	Does the user mainly need to find an existing answer?
Traditional ML	Is the task classification, ranking, forecasting, or anomaly detection with stable labels?
Generative AI	Is useful synthesis or generation required under bounded evaluation?
Human service	Is judgment rare, high-impact, and cheaper to review manually?

Scope design

Define the smallest end-to-end task that creates measurable value. Exclude unsupported users, languages, data sources, channels, high-risk decisions, irreversible actions, and long-tail exceptions until evidence supports expansion.

Use four boundaries:

Capability: what the system can and cannot do.
Authority: what it may do automatically.
Information: what it may access, retain, and reveal.
Responsibility: what remains with the user, operator, and approver.

Requirement structure

Write requirements as observable behavior:

Given [context and permission]
When [user action or event]
The system shall [observable response]
Within [performance constraint]
And shall [evidence / control / audit behavior]
If [failure or uncertainty]
The system shall [refuse / explain / request / escalate / recover]

For each requirement, record priority, rationale, owner, dependency, acceptance, and failure severity.

Traceability

Maintain this chain:

User problem → product goal → requirement → evaluation case → release gate → operating metric

If an item has no upstream problem, question its value. If a high-impact requirement has no evaluation or owner, it is not ready.