Generates an evidence-backed business workflow and an Agent Skill proposal by interviewing operators and product owners. Use when a foreign-trade, manufacturing, engineering, or ecommerce team asks to 梳理流程, 盘点岗位工作, design a repeatable business workflow, or decide which Skills to build before drafting or installing them.
Generates an evidence-backed business workflow and an Agent Skill proposal by interviewing operators and product owners. Use when a foreign-trade, manufacturing, engineering, or ecommerce team asks to 梳理流程, 盘点岗位工作, design a repeatable business workflow, or decide which Skills to build before drafting or installing them.
PM Workflow Planning
Turn one messy business process into a confirmed workflow and a Skill proposal.
This is an experiment: keep the run to planning artifacts. Execute the business
process, edit stable skills/, or promote a draft only after an explicit user
request.
When to use
A business owner can describe the work but not an agent-ready process.
A team needs roles, handoffs, evidence, exceptions, and acceptance criteria
clarified before building Skills.
A live-research workflow needs explicit source-access and browser-evidence
gates.
When NOT to use
Reviewing a finished Skill: use skill-self-check.
Performing the proposed workflow immediately.
Gathering private personal data, bypassing access controls, or storing client
secrets in this repository.
Combining several unrelated processes in one interview.
Check axes
Process scope and intended decision.
Appropriate selection of L/S/V/R business-experience modules.
Measurable target, formula, threshold, owner, and scoring version when L is selected.
Runtime capability and evidence access.
Source fitness (S1–S3) and verification strength (V0–V3).
Roles, handoffs, approvals, and exceptions.
Completion criteria and evidence produced by each step.
Machine-readable run/evidence records when R is selected or an existing log is reused.
Automation boundary and human decision ownership.
Skill proposal contract and promotion gate.
Operating contract
Ask one question at a time and wait for the answer.
For every decision question, give a recommended answer, the reason, and the
main alternative. Let the user confirm or correct it.
Look up discoverable facts with available tools instead of asking the user.
Treat business goals, risk tolerance, ownership, and approval as user
decisions.
Record unsupported or inaccessible facts as unknown — blocker. Keep gaps
open until evidence replaces model memory or “industry common sense.”
Draft the Skill proposal only after the user confirms the workflow summary.
Inputs
One target process.
Company/industry lens.
Desired business outcome or failure to prevent.
Known systems, roles, source restrictions, and approval boundaries.
Done when: the target process is named narrowly enough to interview, or the
first one-question prompt is ready.
Process
Pass 0 — Scope and module-selection gate
Name the process and its intended decision or deliverable.
Read
references/operational-data-contract.md.
Judge L/S/V/R applicability. If the decision is clear from the process, select
modules and explain why. If it is unclear, ask one question with a recommended
selection and wait.
Use N/A — <reason> for skipped modules. Skipping a module with a clear reason
is not a defect and must not be converted into a self-check finding.
Done when: L/S/V/R each has Y, N, or Ask user, with a reason.
Pass 1 — Conditional measurable goal gate
When L is selected, require the user to confirm metric, formula, unit,
direction, owner, L1/L2/L3 thresholds, allowed action, and
score_rule_version. Propose numeric defaults when helpful, then wait for
confirmation. Keep business value L separate from verification strength V.
When L is skipped, write a qualitative completion criterion and the N/A reason.
Done when: L rules are reproducible, or L is explicitly N/A.
Pass 2 — Conditional source and capability gate
When S or V is selected for live web research, read
references/browser-evidence.md. Test and
record whether the current environment can access each required source. Start
Google, LinkedIn, company websites, registries, and browser capabilities in the
not tested state.
When S is selected, create the source register. Assign S1/S2/S3 by claim
fitness, not prestige. Google search remains S1 discovery until the underlying
source is opened.
Done when: selected source/verification modules have their contracts, or
S/V is explicitly N/A.
Pass 3 — Interview one branch at a time
Follow INTERVIEW.md. Resolve 5W2H cells in dependency order:
What and Why.
Who and approval.
Trigger, deadline, and handoffs.
Systems, evidence, exceptions, and volume.
For each question:
Ask one question.
Provide a recommended answer based on confirmed context.
Explain what would change the recommendation.
Wait for confirmation or correction.
Done when: each 5W2H cell is concrete or has an explicit blocker, owner, and
follow-up time.
