omh-deep-interview

OMH Deep Interview — Requirements Specification Through Conversation

When to Use

• The goal is vague, underspecified, or could be interpreted multiple ways
• Before planning (omh-ralplan) or implementation on non-trivial work
• The user says: "deep interview", "requirements", "what should we build", "help me think through this"
• omh-ralplan determines the goal is too ambiguous to plan
• You're unsure what the user actually wants
• Domain unfamiliarity: if the goal requires external knowledge of an unfamiliar domain, suggest running omh-deep-research first to gather context, then resume the interview with the confirmed report as input.

When NOT to Use

• The goal is already crystal clear and bounded
• A confirmed spec already exists in .omh/specs/ for this project
• The user explicitly wants to skip requirements gathering
• Trivial single-file changes where the task is obvious

Prerequisites

• Conversational access to the user (this skill asks questions and needs answers)
• Write access to .omh/ directory for state and spec files

Procedure

Follow these phases in order. The skill operates through conversation with the user and file writes for state and spec output.

Phase 0: Check for Existing State

Before starting a new interview:

1. Enumerate active interviews:

   listed = omh_state(action="list_instances", mode="interview")
   

   Each entry carries instance_id (the interview id) and active flag. If omh_state is unavailable, glob .omh/state/interview--*.json manually.
2. If any active interview exists, tell the user: "There's an active interview for '{project_name}' (id={id}). Resume, start fresh, or abandon it?"
3. If resuming: omh_state(action="read", mode="interview", instance_id="{id}") — read round summaries to reconstruct context
4. If abandoning: omh_state(action="write", mode="interview", instance_id="{id}", data={...status: "abandoned"}), then proceed to Phase 1 with a NEW id
5. If no active state found: proceed to Phase 1
6. Concurrent interviews on different projects are permitted; do not block.
7. Check for existing research context (omh-deep-research sentinel): if any .omh/research/*-report.md exists with frontmatter status: confirmed, mention it to the user as available context for the interview (e.g., "I see a confirmed research report on '{topic}' at {path} — want me to fold that in as background?"). Do NOT auto-load it; the user decides.

Phase 1: Opening

Start the interview with two questions:

1. Project description: "Describe what you want to build in 2-3 sentences. What's the core idea?"
2. Greenfield or brownfield: "Is this a new project from scratch, or are you working within an existing codebase or system?"

Then:

• Generate an interview ID: di-{YYYYMMDD}-{short_random} (e.g., di-20260407-x7k)
• Ask the user for a short project name (for filenames)
• Create the state file at .omh/state/interview--{id}.json (engine derives this from instance_id="{id}") with:
  • All coverage dimensions set to HIGH
  • type set to greenfield or brownfield based on user's answer
  • existing_context coverage set to N/A for greenfield projects
  • Round 0 summary capturing the opening answers
• Tell the user: "Got it. I'll ask up to 5 rounds of questions to clarify requirements. You can say 'enough' at any point if you feel we've covered what's needed."

Phase 2: Interview Loop

Run up to 5 rounds (extensible to 10 if user requests). Each round:

Step 1 — Select dimension to probe

Compare coverage bins across dimensions. Target the dimension with the highest ambiguity. When multiple dimensions share the same bin, use weights to break ties:

• Greenfield: Goal (0.40) > Constraints (0.30) = Success Criteria (0.30)
• Brownfield: Goal (0.35) > Constraints (0.25) = Success Criteria (0.25) > Existing Context (0.15)

Load references/scoring-rubric.md for detailed bin definitions and examples.

Step 2 — Ask the question

Ask ONE primary question targeting the selected dimension. Make it specific and grounded in what the user has already told you. Don't ask generic questions — reference their earlier answers.

Good: "You mentioned this is a CLI tool for personal use. What happens when you run it with no tasks configured — should it create a default list, show an error, or something else?"

Bad: "What are your constraints?"

Allow 1-2 brief follow-ups if the user's answer is unclear, contradictory, or raises new questions. Don't force follow-ups if the answer is clear.

Step 3 — Adaptive questioning (if stuck)

If a dimension has been targeted for 2 or more consecutive rounds without moving from its current bin, change your approach:

• Try asking from a completely different angle
• Propose a concrete example and ask "Is this right, or what would you change?"
• Ask what's blocking clarity: "What makes this hard to pin down?"
• Ask the user to define a specific term they've been using ambiguously

Step 4 — Update coverage

After processing the user's answer, re-assess the coverage bin for the targeted dimension. Use the rubric in references/scoring-rubric.md. Be conservative — when in doubt, keep the bin at its current level rather than prematurely lowering it.

Step 5 — Update state

Update the interview state (increment round, add summary, update coverage):

omh_state(action="write", mode="interview", instance_id="{id}", data={...updated state with new round...})

Each round summary: max ~200 words — capture what was learned, not the full exchange.

Step 6 — Present coverage and ask to continue

Show the user where things stand:

Coverage after round {N}:
  Goal:             [MEDIUM]  →  Some clarity, but scope needs bounding
  Constraints:      [LOW]     →  Mostly clear, data format TBD
  Success Criteria: [HIGH]    →  Need testable criteria
  Existing Context: [N/A]     →  Greenfield project

Then ask: "Want to continue refining, or is this enough to work with?"

• If the user says continue: proceed to next round
• If the user says enough/done/that's it: proceed to Phase 3
• If round 5 reached: "We've done 5 rounds. Want to continue (up to 10 more), or shall I generate the spec from what we have?"
• If round 10 reached: proceed to Phase 3 automatically

The user always controls exit. Never auto-terminate based on coverage scores.

Phase 3: Spec Generation

When the user confirms exit (or max rounds reached):

1. Load templates/spec-template.md
2. Synthesize a specification from the accumulated round summaries in the state file
3. Fill in every section of the template:
   • Goal: Synthesize from rounds that targeted the goal dimension
   • Constraints: Synthesize from constraint-focused rounds
   • Success Criteria: Must be specific and testable — if a criterion can't be verified, flag it in Open Questions instead
   • Existing Context: Only for brownfield; synthesize from context-focused rounds
   • Assumptions: Things the user stated or implied that haven't been validated
   • Open Questions: Anything with coverage still at HIGH or MEDIUM — these are unresolved
4. Set YAML frontmatter:
   • status: draft
   • coverage: current bin values
   • rounds_completed: total rounds done
5. Write to .omh/specs/{project-name}-spec.md
6. Update state: spec_file = path to spec

Phase 4: Confirmation

Display the full draft spec to the user and ask:

"Here's the specification I've drafted. Please review it. You can:

• Confirm — I'll finalize it and downstream skills (ralplan, autopilot) will use it
• Request changes — Tell me what to adjust, and I'll ask follow-up questions
• Abandon — Discard the spec and interview"

Handle each response:

On confirm:

• Update spec frontmatter: status: confirmed
• Update state: status: confirmed
• Tell the user: "Spec confirmed and saved to {path}. You can now use omh-ralplan to create an implementation plan from this spec."

On request changes:

• Ask targeted follow-up questions about the specific sections the user wants changed
• Regenerate the spec
• Return to the confirmation prompt

On abandon:

• Delete the spec file
• Update state: status: abandoned
• Tell the user the interview has been discarded

Only specs with status: confirmed are considered valid by downstream skills.

Phase 5: Logging

Throughout the interview, write structured events to .omh/logs/interview-{id}.log:

2026-04-07T06:30:00Z STARTED interview_id=di-20260407-x7k project=my-project type=greenfield
2026-04-07T06:31:15Z ROUND round=1 dimension=goal coverage_change=goal:HIGH→MEDIUM
2026-04-07T06:33:42Z ROUND round=2 dimension=constraints coverage_change=constraints:HIGH→LOW
2026-04-07T06:35:00Z USER_EXIT round=3 reason=user_confirmed
2026-04-07T06:35:30Z SPEC_GENERATED path=.omh/specs/my-project-spec.md status=draft
2026-04-07T06:36:00Z SPEC_CONFIRMED path=.omh/specs/my-project-spec.md

Log events and decisions only — NOT conversation content.

Sentinel Convention

Downstream skills detect completed interviews by checking for files matching .omh/specs/*-spec.md with status: confirmed in the YAML frontmatter.

• omh-ralplan: If a confirmed spec exists, use it as the goal/specification input to the Planner subagent instead of asking the user to describe the goal.
• omh-autopilot: If a confirmed spec exists, skip Phase 0 (requirements) entirely.

State Management

State file: .omh/state/interview--{id}.json (engine path; pass instance_id="{id}" to omh_state)

See references/state-schema.md for the full schema.

Key rules:

• Each interview is a separate instance keyed by its id; concurrent interviews on different projects are permitted
• State stores round summaries (max ~200 words each), NOT full transcripts
• State is preserved after completion (for audit trail)
• Resumability is inherently lossy — round summaries help reconstruct context but lose nuance from the original conversation

Pitfalls

• Never auto-terminate based on scores. Coverage bins are advisory heuristics for question targeting. The user always decides when they're done.
• Don't store full transcripts. Round summaries keep state compact and respect context window limits on resume. Accept that resumability is approximate, not exact.
• Ask about brownfield, don't auto-detect. Checking for package.json etc. is unreliable and presumptuous. Let the user tell you.
• One active interview per project. Use a fresh instance_id per interview. If an active interview already exists for the same project, offer to resume or abandon it before starting a parallel one with the same id.
• Spec must be confirmed. Draft specs are not valid input for downstream skills. The user must explicitly confirm.
• Don't ask generic questions. Reference the user's earlier answers. "What are your constraints?" is lazy. "You mentioned this is for personal use on Linux — are there other platforms it needs to work on?" is useful.
• Be conservative with coverage assessment. When in doubt, keep the bin at its current level. It's better to ask one more question than to prematurely declare a dimension CLEAR.