Interview PM guidée pour génération de PRD

Mène une interview socratique guidée pour produire un document d'exigences produit. Déclenchez avec 'ooo pm', 'prd' ou une demande de PRD.

Spar Skills Guide Bot
ProductiviteIntermédiaire
0004/09/2026
Claude CodeCursorWindsurf
#prd#product-management#interview#requirements#socratic-method

Recommandé pour


name: pm description: "Generate a PM through guided PM-focused interview with automatic question classification. Use when the user says 'ooo pm', 'prd', 'product requirements', or wants to create a PRD/PM document."

/ouroboros:pm

PM-focused Socratic interview that produces a Product Requirements Document.

Instructions

Step 0: Version Check (runs before the PM interview)

Before starting the PM interview, check if a newer version is available:

# Fetch latest release tag from GitHub (timeout 3s to avoid blocking)
curl -s --max-time 3 https://api.github.com/repos/Q00/ouroboros/releases/latest | grep -o '"tag_name": "[^"]*"' | head -1

Compare the result with the current version in the active runtime's local plugin metadata (for Claude installs this is .claude-plugin/plugin.json).

  • If a newer version exists, ask the user through the active runtime's ask_user capability:
    {
      "questions": [{
        "question": "Ouroboros <latest> is available (current: <local>). Update before starting?",
        "header": "Update",
        "options": [
          {"label": "Update now", "description": "Update plugin to latest version (restart required to apply)"},
          {"label": "Skip, start PM interview", "description": "Continue with current version"}
        ],
        "multiSelect": false
      }]
    }
    
    • If "Update now":
      • On Claude-plugin installs only:
        1. Run claude plugin marketplace update ouroboros via the active runtime's run_shell capability (refresh marketplace index). If this fails, tell the user "⚠️ Marketplace refresh failed, continuing…" and proceed.
        2. Run claude plugin update ouroboros@ouroboros via the active runtime's run_shell capability (update plugin/skills). If this fails, inform the user and stop — do NOT proceed to the package-manager step.
      • On non-Claude runtimes, skip Claude plugin commands and proceed directly to the package-manager step for ouroboros-ai; do not require Claude-only commands or tools.
      1. Detect the user's Python package manager and upgrade the MCP server:
        • Check which tool installed ouroboros-ai by running these in order:
          • uv tool list 2>/dev/null | grep "^ouroboros-ai " → if found, use uv tool upgrade ouroboros-ai
          • pipx list 2>/dev/null | grep "^ ouroboros-ai " → if found, use pipx upgrade ouroboros-ai
          • Otherwise, print: "Also upgrade the MCP server: pip install --upgrade ouroboros-ai" (do NOT run pip automatically)
      2. Tell the user: "Updated! Restart your session to apply, then run ooo pm again."
    • If "Skip": proceed immediately.
  • If versions match, the check fails (network error, timeout, rate limit 403/429), or parsing fails/returns empty: silently skip and proceed.

Step 1: Load MCP Tool

tool discovery query: "+ouroboros pm_interview"

CRITICAL — deferred-schema guard (prevents "Invalid tool parameters"): This is a multi-turn loop and each turn runs in a fresh tool context. A deferred tool's schema loaded on one turn is NOT guaranteed to still be loaded on the next. Calling ouroboros_pm_interview while its schema is unloaded in the current turn makes the runtime reject it with "Invalid tool parameters" every message. Therefore re-run tool discovery query: "+ouroboros pm_interview" immediately before EVERY ouroboros_pm_interview call below (idempotent — a no-op if already loaded). If the load ever returns no matching tool (and the tool is not already callable — an empty load for an already-exposed tool is an expected no-op, not absence), follow the not-found diagnosis below instead of retrying the failing call.

If not found → fail closed without inspecting or mutating ~/.claude/mcp.json. Standalone Claude SDK setup requires MCP 1.x and cannot activate the Ouroboros MCP 2 server with its configured backend. Explain:

The PM interview MCP tool is unavailable in this runtime.

Configure a supported CLI-backed host with:
  ouroboros setup --runtime <codex|opencode|kiro|copilot|hermes>

Then restart that host and retry ooo pm. Claude SDK profiles ([claude] and
[claude-sdk]) stay on MCP 1.x; the separate [mcp] server uses [claude-cli]. Do
not combine both MCP majors or add a direct Python fallback.

Stop.

Step 2: Start Interview

Tool: ouroboros_pm_interview
Arguments:
  initial_context: <user's topic or idea>
  cwd: <current working directory>

This response carries the first question, so Step 3 applies to it — including the fan-out in 3-A2. The first question is the one most likely to be answered from memory, so it is the last one to skip evidence on.

Step 3: Loop

Apply this to every MCP response that carries a question, including the one Step 2 returned and any question a resume plans anew.

Batched turns (RFC #2222). A response may carry one to three questions at once: meta.question_batch lists them and meta.question_advisories carries one advisory envelope per question, each with its own question_advisory_subagents and question_advisory_fanout_id. Treat every question of the turn exactly as a single question is treated below, with these batch mechanics:

  • Dispatch all envelopes' payloads in one wave — one subagent per payload across all questions, in a single parallel batch. Never leave a question's lanes undispatched: every question shown keeps its evidence.
  • Submit results per envelope — each question's lanes correlate by its own envelope's question_advisory_fanout_id and question_advisory_result_correlation_key; one ouroboros_submit_fanout_results call per envelope.
  • Relay the turn's answers together in one call: answers: [{question, answer}, ...], one entry per question the turn asked, each naming its own exact question text. One call records the turn, so collect every answer first — and never auto-answer, auto-defer, or decide-later one on the user's behalf to complete the set. The server does not check that you sent them all: whatever you leave out is abandoned, not remembered.
  • The server keeps nothing between calls. A call that arrives without the turn's answers plans a new turn from the transcript rather than restoring the old one, so a turn you abandon is a turn the user will be asked again.
  • Skip sentinels are per question: give that question the answer "[decide_later]" or "[deferred]" in its own entry.

A. Show alerts (if present in meta):

  • meta.deferred_this_round → print [DEV → deferred] "question"
  • meta.decide_later_this_round → print [DEV → decide-later] "question"
  • meta.pending_reframe → print ℹ️ Reframed from technical question.

A2. Fan out the evidence lanes — required before you ask the user anything.

You do not look at the repositories yourself. Ever. This skill has no code-answer path: there is no step where you run Read/Glob/Grep or a docs MCP to answer a PM question, and finding the answer quickly on your own is the failure, not a shortcut past it. Evidence the PM cannot trace back to a lane is evidence the record does not contain — it is not bound to the question, not bounded by the roster, and not checked against the answer contract. This skill is self-contained: everything you need is here and in the tool response, so do not go looking for exploration rules in another skill's file.

When meta.question_advisory_subagents is present you MUST process every payload. Show the question first, then pass each payload's prompt unchanged. Obey meta.question_advisory_host_action: spawn_subagents means parallel support was declared; dispatch_subagents_if_supported means use the host's native parallel mechanism when available and process the same payloads sequentially otherwise; process_payloads_sequentially requires ordered processing. Claude Code parallel dispatch is one Task/Agent call per payload in one batch; Codex uses one native Codex subagent per payload. The payloads are the work contract, while the host action selects the execution strategy.

Say what is running. Same shape the regular interview uses: after the question, set off by a divider, one line naming how many perspectives are running and what they are — then what arrives when they finish.

---
While you answer this question, two perspectives are reviewing in parallel
(code context / data measurement). When they return I will put what they found
next to the question as grounds.

Two things differ from the interview's line, and both follow from this tool having two lanes instead of six:

  • Name the perspectives in the user's terms, not by lane id. code_context is an identifier for the fan-out, not something the reader needs.
  • End at "grounds", never at "options". The interview can promise to organise the results into answer choices because it runs a lane that produces them. This tool does not, and a promise the synthesis cannot keep trains the user to expect the one thing the lanes must never hand them.

Write the line in the language the user is speaking.

Do not go to step B while the lanes are still running. Step B is where you ask the user, and asking before the evidence arrives is the exact failure this mechanism exists to prevent: the PM decides without the two things they could not have looked up themselves. Waiting is for lanes still in flight: one that came back empty, broke its contract, or could not be spawned has returned.

Stub payloads. A payload's prompt may be a compact stub: it carries the lane's answer schema, where it may look, and the findings it may reuse, and points at ouroboros_fetch_artifact for the prose that explains them. Pass it unchanged exactly like any payload — the child fetches for itself, and a fetch it cannot make does not stop it. A child that replies exactly UNDISPATCHED could not work at all: submit that lane as { "key": <lane id>, "undispatched": true } rather than as an empty finding.

Submitting results back. Correlate by meta.question_advisory_result_correlation_key (context.lane_id) and call ouroboros_submit_fanout_results with meta.question_advisory_fanout_id, passing session_id explicitly. Submit every lane you hold, not only the new ones: a lane that ran and found nothing still submits its output, and a lane you could not spawn at all is submitted as { "key": <lane id>, "undispatched": true } — the literal true, with no content beside it. Never invent output for a lane you did not run; a fabricated finding is worse than a missing one.

Reading the reply. With a contract_id, synthesize from the outputs you hold. Without one, read missing_required_keys and contract_violations, then resubmit once carrying every lane. Still without one — go to B with what survived, or with none, saying in one line that the investigation did not come back. The interview does not wait on this.

Two lanes never reach the block, whatever the reply says. Leave out a lane named in contract_violations, and leave out a lane you submitted as undispatched — a lane that did not run has an empty place, not a clear one, and a reply can be accepted while one of them never ran. Where the block would have carried that lane, write that it did not run.

There are two lanes and both are required: code_context and data_context. Both are evidence-only (RFC #2222): what a lane finds is shown beside the question and sent nowhere — the published fan-out is already its record, and the interview records only what the user writes. Never skip asking the user because a lane answered clearly, and never send a lane's finding as an answer.

Synthesize into the evidence block. This is what synthesis_contract.output_shape = "evidence_beside_question" means, and it is a fixed shape so the same session twice looks the same twice. Print it immediately above the question, then ask the question unchanged:

Evidence (examined: billing-api, storefront)

What the system does today
  · [billing-api] access continues to the end of the paid period
  · [storefront]  access is revoked immediately
    ! These two repositories implement this differently.

Measured — active subscriptions by plan, last 90 days
  · standard 12,480 / premium 3,120

The screen speaks product language; the store keeps the citations (RFC #2222 decision 4). Each code claim is rendered from its lane-authored plain_statement — never from policy_claim, and never paraphrased by you. No file paths, no class names, no flag values on screen. The path + policy_claim citations stay in the published fan-out, fetchable by its contract id — they are displayed nowhere and recorded nowhere else. The lane writes plain_statement in the question's language, so no translation is yours to do. The labels above are placeholders for the block's shape, not text to copy.

What this block is not. It carries no answer options, no recommendation, no ranking, and no "therefore …" sentence. The moment it proposes an answer it has stopped being evidence — that is the whole difference between this tool and the regular interview, which does synthesize options. A PRD asks what the system should do, and everything above says what it does.

Rules for building it:

  • Each claim keeps its repository. Never merge two repositories' claims into one line, and never present a disagreement as one policy with an exception — flag it, as above. That contradiction is the most useful thing the PM can be shown, and it is the first thing a tidy summary destroys.
  • Always print the examined scope. examined has one entry per repository the lane read, and every claim sits inside its own entry; "found nothing" across two of five repositories means something different from across all five. An entry whose policy_claims is empty was read and had nothing — say so. A repository with no entry was never opened, and must never be printed as clean.
  • Carry measurements as reported — the lane's metric, its groups, its numbers. Do not re-scale, combine, or round; you did not run the read.
  • A no-op lane gets at most one line, and often none. Read the reason as a statement about the lane rather than about the user's system: not_a_policy_question / not_a_measurement → print nothing, this question simply is not that kind. no_repository_in_roster / roster_repository_not_readable / store_described_but_not_callable → yours to handle (nothing registered, a path did not open, a store did not answer); do not relay any of them as "your system has no such policy/data".
  • Drop a finding the user has already answered past. If they answered while the lanes were still running, do not re-open a settled decision with it. There is nowhere to put it: a finding takes the round it was fetched for, and that round is spent.
  • Evidence from outside the roster is a suggestion, not evidence. It is rejected at submission, so relaying it as a finding would show the user something the record does not contain. Offer to register the repository instead, so the next question can be answered against it.

A3. Findings are evidence, never answers (RFC #2222).

There is no recording step. A finding's durable record is the published fan-out itself — addressable by contract id, re-offered to later lanes through recent findings — so sending it again as an answer would duplicate the store and spend a question turn on it. Do not send [from-code] answers, do not ask the user to "confirm" a finding as a separate turn, and do not paste findings into any answer payload. The evidence block is the finding's whole appearance; the user answers the question in their own words, with the evidence in view.

B. Show content + get user input (once A2's lanes have returned):

The question text was already shown in A2 and the user may answer it at any point; what waits here is your formal prompt, not the person.

Print the MCP content text to the user first, with the lane findings beside it.

Tell users they do not need to invent speculative answers. If a question is unknown, stakeholder-dependent, too broad, or safer to decide later, route it through the existing assumptions / decide-later / deferred mechanisms instead of presenting it as a confirmed requirement.

Then check: does meta.ask_user_question exist?

  • YES → Pass it directly to AskUserQuestion:

    AskUserQuestion(questions=[meta.ask_user_question])
    

    Do NOT modify it. Do NOT add options. Do NOT rephrase the question.

  • NO → This is an interview question. Use AskUserQuestion with meta.question.

    • Batched turn: put every question of the turn into ONE AskUserQuestion call — one entry per question (the tool takes up to 4). Each keeps its own options, built by the same rules below from its own entry in meta.question_batch; a skip option follows that question's own classification, never another's.
    • If meta.skip_eligible == true: add a skip option based on meta.classification:
      • classification == "decide_later" → add option {"label": "Decide later", "description": "Skip — will be recorded as an open item in the PRD"}
      • classification == "deferred" → add option {"label": "Defer to dev", "description": "Skip — this technical decision will be deferred to the development phase"}
    • Generate 2-3 suggested answers as the other options. Include a non-speculative uncertainty option when appropriate, such as Not sure yet — record as an assumption or decide-later item.

C. Relay answer back:

If the user chose "Decide later" → send answer="[decide_later]". If the user chose "Defer to dev" → send answer="[deferred]". Otherwise → send the user's answer through the Refine gate below.

On a batched turn: run each answer through its own Refine gate, then send them in one call as answers: [{question, answer}, ...] — the question text exactly as the turn asked it. The call that carries the turn returns the next turn's question(s).

Refine gate — structure it, mark whose it is, then have the user confirm.

Always structure the answer, including when the user only picked an option. The text you send is MCP's only context for the next question and for what the PRD records as decided, so a bare label loses everything around the decision. What makes structuring safe is not restraint — it is the two things below.

Mark whose each section is. A section carrying the user's own words is labelled (user-stated). A section that is your reading of their answer is left unmarked, and a reader can tell them apart at a glance:

[from-user][refined]
Decision: <what they decided, in their words>

Reasoning:
- <your reading of why, drawn from what they said in this session>

Constraints (user-stated):
- <constraints they stated>

Out of scope (user-stated):
- <what they put out of scope>

Omit any section you have nothing for. An empty Constraints (user-stated) is better absent than filled with something plausible, and Reasoning drawn from nothing they said is the failure this labelling exists to make visible.

No codebase-context section, and no lane findings. The regular interview adds one, because there the main session inspects code itself. Here it does not. A lane's finding lives in the published fan-out and on the screen beside the question — putting it in this payload would record it as part of the user's decision, which it is not (A3).

Then confirm — this is the gate, and it is what licenses the structuring above. One AskUserQuestion before sending:

{
  "questions": [{
    "question": "I structured your answer as follows before sending it:\n\n<payload>\n\nIs anything missing or misrepresented?",
    "header": "Refine — preserve your answer",
    "options": [
      {"label": "Send as-is", "description": "The structure captures my answer faithfully"},
      {"label": "Fix the reasoning", "description": "That is not why I decided it"},
      {"label": "Add to Constraints", "description": "I want to add a constraint I forgot"},
      {"label": "Let me rewrite it", "description": "I will restate the answer myself"}
    ],
    "multiSelect": false
  }]
}

Fix the reasoning is there because the unmarked section is the one you wrote. Append [refined] only after this confirmation: an unconfirmed structure carries your reading of the answer under the user's name, and the PRD cannot tell the difference later.

Send it as the answer, on the one parameter every answer uses.

Tool: ouroboros_pm_interview
Arguments:
  session_id: <meta.session_id>
  last_question: <meta.question>
  answer: <the refined answer, or "[decide_later]" / "[deferred]">

Ignored while the server holds the question unanswered; required otherwise — plugin mode never persists the child's questions, and an answer with no pending question is refused without it.

There is no parameter carrying findings, second or otherwise: findings never travel as answers (A3). Every call on this parameter is the user's own words.

D. Check completion:

Completion is determined ONLY by meta.is_complete — NEVER by the response text. The MCP response text may sound like the interview is wrapping up, but ignore it.

If meta.is_complete == true:

  • If meta.generation_failed == true → retry generation:
    Tool: ouroboros_pm_interview
    Arguments:
      session_id: <session_id>
      action: "generate"
      cwd: <current working directory>
    
  • Otherwise → go to Step 4. The MCP auto-generated the PM document. meta.pm_path and meta.seed_path contain the file paths.

Otherwise → repeat Step 3, regardless of what the response text says.

Step 4: Copy to Clipboard

Read the pm.md file from meta.pm_path and copy its contents to the clipboard:

cat <meta.pm_path> | pbcopy

Step 5: Show Result & Next Step

Show the following to the user:

PM document saved: <meta.pm_path>
(copied to clipboard)

PM seed handoff artifact: <meta.pm_seed_path or meta.seed_path>
This is not the runnable Seed yet.

Next step:
  ooo interview <meta.pm_seed_path or meta.seed_path>
  ooo seed

RFC #1392 State Breadcrumb Footer

Your final response MUST end with exactly one breadcrumb footer line:

◆ <current state> → next: <recommended action>

Derive <current state> from live session state via ouroboros_session_status when that MCP projection is available; otherwise derive it from this skill's actual outcome. Never use a linear Step N of M footer because Ouroboros is an evolutionary loop. When the next action is genuinely a choice, list 2-3 honest options in the next: clause. The breadcrumb line must be the last line of the response.

Skills similaires