/eneo-discuss

/eneo-discuss

Run a structured Socratic interview over a draft PRD and raise or lower confidence until it crosses the 90/70 gate. Deep lane only — Standard and Fast lanes skip this step.

Pre-flight

1. Read .claude/state/current-task.json. If lane != "deep", print:

   ✗ /eneo-discuss is Deep-lane only. Current lane: <lane>.
     Fix: run /eneo-start to proceed, or /eneo-new to re-classify.
   

   and exit.

2. Recommend plan mode. If the developer is not in plan mode, print once:

   Hint: plan mode keeps the interview exploratory. Press Shift+Tab to enter.
   

   Do not block; continue.

3. Determine slug — $ARGUMENTS takes precedence, otherwise current-task.json.slug. If neither is present, print Next: /eneo-new "<description>" and exit.

Interview script (Pocock, verbatim prompts)

Stream one section at a time; confirm between sections.

1. Problem statement. "Describe the problem in detail. Who feels the pain; how do you know it's happening; what would success look like." Verify the PRD's Problem statement against the answer. If they diverge, edit the PRD.
2. Repo assertions. Dispatch a read-only Explore subagent (fresh context) with the PRD's Module sketch, Success criteria, and Acceptance criteria. Instruct it to return CONFIRMED|<file:line> or CONTRADICTED|<file:line> — <reason> for each claim. Print each return in real time.
3. Alternatives. "Which options did you consider? Why reject each?" Record.
4. Scope. "Hammer out the exact scope. What is explicitly out of scope?" Update ## Out of scope in the PRD accordingly.

Confidence gate (SuperClaude 90/70)

Compute a confidence score 0–100 based on:

• Problem statement has a named actor + measurable pain signal (+20)
• Success criteria are numeric or command-executable (Section B BAD/GOOD rule) (+20)
• Every CONTRADICTED from the Explore subagent has been reconciled in the PRD (+20)
• Out of scope is populated (+10)
• Open questions labeled (not silently dropped) (+10)
• At least one security/authz user story when auth surface touched (+10)
• Swedish-language UX / a11y story when frontend touched (+10)

Then:

• ≥ 90% — print ✓ Confidence <N>% — ready to plan. Next: /eneo-plan and exit.
• 70–89% — present 2–3 concrete alternatives / clarifying questions and ask the developer to answer. Re-score; loop.
• < 70% — block with specific missing items. Print Confidence <N>% — not ready. Fill the gaps above and re-run. Exit 0; do not proceed.

DX rules

• Stream each interview section and each Explore CONFIRMED|CONTRADICTED return as it happens; never batch.
• Use AskUserQuestion only between the interview sections when the developer needs to choose between alternatives. Do not prompt for anything else.
• The final line is always Next: <command> inferred from the gate result.