Imported from Rune-kit/rune (skills/ba/SKILL.md). Install upstream with npx skills add Rune-kit/rune --skill ba. Copyright stays with the author.

ba

Purpose

Business Analyst agent — the ROOT FIX for "Claude works a lot but produces nothing." BA forces deep understanding of WHAT to build before any code is written. It asks probing questions, identifies hidden requirements, maps stakeholders, defines scope boundaries, and produces a structured Requirements Document.

Wrong requirements shipped correctly is the most expensive bug. BA's job is to prevent it — measure clarity (Step 2.5), measure completeness (Step 3.5), and measure cross-dimension consistency (Step 3.6) before handoff.

Goal-first (advisory, 2026). The Requirements Document is the goal spec current models want up front. Where the platform has a native goal/outcome command — Claude Code /goal, Managed Agents Outcomes (rubric) — seed it from this document's scope boundaries + acceptance criteria so the whole run stays anchored to the agreed WHAT.

Triggers

• Called by cook Phase 1 when task is product-oriented (not a simple bug fix)
• Called by scaffold Phase 1 before any project generation
• /rune ba <requirement> — manual invocation
• Auto-trigger: when user description is > 50 words OR contains business terms (users, revenue, workflow, integration)

Calls (outbound)

• scout (L2): scan existing codebase for context
• research (L3): look up similar products, APIs, integrations
• plan (L2): hand off Requirements Document for implementation planning
• brainstorm (L2): when multiple approaches exist for a requirement
• design (L2): when requirements include UI/UX components — hand off visual requirements

Called By (inbound)

• cook (L1): before Phase 2 PLAN, when task is non-trivial
• scaffold (L1): Phase 1, before any project generation
• plan (L2): when plan receives vague requirements
• brainstorm (L2): standalone ideation that picked an approach for a new feature with no spec — hands the chosen approach to ba for requirements before plan
• mcp-builder (L2): requirements elicitation before MCP server design
• User: /rune ba direct invocation

Cross-Hub Connections

• ba → plan — ba produces requirements, plan produces implementation steps
• ba → brainstorm — ba calls brainstorm when multiple requirement approaches exist
• ba ↔ cook — cook calls ba for non-trivial tasks, ba feeds requirements into cook's pipeline
• ba → scaffold — scaffold requires ba output before project generation

Executable Steps

Step 1 — Intake & Classify

Read the user's request. Classify the requirement type:

If Bug Fix → skip BA, route to cook/debug directly. If Refactor → light version (Step 1 + Step 4 only). Skip Steps 2, 2.5, 3, 5, 6.

If existing codebase → invoke rune:scout for context before proceeding.

Step 1.4 — Synthesis Trigger Check

Before proceeding to elicitation, check whether the requirements are already in context. Re-asking what the user already told you is the second-most expensive bug.

Activate Synthesis Mode instead of standard elicitation if ANY of:

In Synthesis Mode: extract answers from existing context, draft the Requirements Document with source citations for every section, then confirm rather than re-interview. Ask follow-ups ONLY on the 1-2 dimensions with genuine gaps. Skip steps 2, 2.5 if all 5 dimensions are filled or partial-but-acceptable.

Workflow detail + anti-patterns: references/synthesis-mode.md.

Step 1.5 — Out-of-Scope Match Check (READ)

Before any elicitation, check whether the request matches a concept previously rejected.

1. Glob .out-of-scope/*.md — if directory absent, skip silently.
2. For each file, parse YAML frontmatter (concept, aliases).
3. Build a token map (lowercased, split on - and whitespace).
4. Tokenize the user's request the same way.
5. Compute lexical overlap per concept; keep the top match's confidence (0.0–1.0).

Action by confidence:

Emit outofscope.match signal with {concept, confidence, verdict} so downstream skills (cook, plan) inherit the context.

If verdict is exact-match AND user says "yes I still want it" → record their override reason in the Requirements Document ## Risks section AND mark priority_to_revisit: high in the existing .out-of-scope/<slug>.md (do NOT delete the file). The override forces the candidate up the revisit ladder; it doesn't erase the prior decision.

If verdict is exact-match AND user accepts the prior rejection → end the BA session with a one-line summary referencing the file. No further questions.

Format reference: references/out-of-scope-format.md.

Step 1.6 — Mid-Elicitation Reject WRITE Path

If the user explicitly rejects the feature at any point during elicitation (Steps 2-3) — common phrases: "scrap it", "actually nah, don't build this", "we won't do this", "kill the feature", "drop it" — STOP elicitation and write a .out-of-scope/<slug>.md record before ending the session.

Without this WRITE path, oral rejections vanish — the next session re-asks the same questions and the user has to re-reject. Step 1.5 (READ) only catches matches against existing files; Step 1.6 (WRITE) is what produces those files in the first place.

Procedure:

1. Confirm rejection is durable, not deferral. Ask one clarifier:

   "Just to record this correctly: is this out of scope (project doesn't want this), or deferred (not now but maybe later)? Out-of-scope gets recorded so we don't re-litigate; deferred goes to backlog instead."

   • If deferred → route to backlog (no .out-of-scope/ write), end session with a one-line note
   • If out-of-scope → continue to step 2

2. Capture the durable reason. Ask:

   "What's the reason this is out of scope? (project scope, technical constraint, strategic decision — not a temporary circumstance)"

   If the user gives a temporary reason ("we're busy"), reframe: "That's a deferral — should I route to backlog instead?"

3. Generate slug (kebab-case, ≤40 chars, recognizable without opening the file).

4. Lexical-similarity check: Glob .out-of-scope/*.md, parse each frontmatter's concept + aliases, compute overlap. If any existing concept has ≥0.7 overlap → APPEND to that file's prior_requests list and mark rejected_by: ba for this round. Do NOT create a duplicate.

5. Write the file using the format in references/out-of-scope-format.md:

   • YAML frontmatter (concept, aliases, decision: rejected, rejected_at, rejected_by: ba, prior_requests, optional revisit_if)
   • Markdown body: concept name, "Why out of scope" (substantive reasoning from step 2), "What would change our mind" (if user volunteered signals)

6. Emit outofscope.recorded signal carrying {slug, rejected_by: ba, prior_requests_count} so downstream skills know a new rejection landed.

7. End BA session with one-line summary:

   "Recorded as out of scope in .out-of-scope/<slug>.md. Future similar requests will surface this. Override anytime by editing the file."

When NOT to write:

• User merely defers ("not now") → backlog, not .out-of-scope/
• User rejects a single requirement within a larger feature → adjust requirements doc Boundaries section, don't write a whole rejection file (the feature is still in scope)
• Bug rejections (already fixed, not reproducible) → not BA's job; route to incident or close the issue
• The match was already exact (≥0.8) and Step 1.5 surfaced it — user accepting the prior rejection just appends to prior_requests of the existing file (handled in Step 1.5 path)

Step 2.0 — Explore-First Pre-Check (HARD-GATE)

Before emitting ANY of the 5 elicitation questions, run the 4-item pre-check on each intended question:

1. Is the answer in package.json / pyproject.toml / Cargo.toml / go.mod / pom.xml?
2. Is the answer in README.md / CLAUDE.md / docs/?
3. Is it inferable from file extensions, directory structure, or config files?
4. Has the user answered it earlier in this conversation?

The gate is "tried to infer" — not "must succeed in inferring." If the file genuinely doesn't have the answer, the attempt itself is the gate.

Cache inferred answers in the requirements doc:

**Inferred from package.json**: TypeScript 5.4, Next.js 14.2, React 18.3
**Inferred from .github/workflows/**: CI runs on PRs targeting main

Worked examples + edge cases: references/explore-first.md.

Step 2 — Requirement Elicitation (the "5 Questions")

Ask exactly 5 probing questions, ONE AT A TIME (not all at once):

1. WHO — "Who is the end user? What's their technical level? What are they doing right before and after using this feature?"
2. WHAT — "What specific outcome do they need? What does 'done' look like from the user's perspective?"
3. WHY — "Why do they need this? What problem does this solve? What happens if we don't build it?"
4. BOUNDARIES — "What should this NOT do? What's explicitly out of scope?"
5. CONSTRAINTS — "Any technical constraints? (existing APIs, performance requirements, security needs, deadlines)"

Question Discipline (MANDATORY)

Every question the user answers burns attention you don't get back. Protect it.

1. Max 5 questions total across the whole BA session. If you find yourself wanting a 6th, the answer is in the first 5 or you're stalling — re-read, don't re-ask.
2. Prefer yes/no or multiple-choice over open-ended. An open-ended question is a last resort when no reasonable option set exists.
   • BAD: "What auth strategy do you want?"
   • GOOD: "Auth: (a) email+password with JWT, (b) OAuth (Google/GitHub), (c) magic link, (d) I'll decide — pick one."
3. Never ask what you can infer. If the answer is in the repo, the user's message, or the classification from Step 1 — don't ask it.
   • Wrong stack? → read package.json, don't ask.
   • Wrong audience? → check the README, don't ask.
   • Wrong framework? → check config files, don't ask.
4. Cache the answer. Write each Q→A pair into the Requirements Document verbatim. If the user restarts the BA session on the same feature, reuse the cached answers — never re-ask what was already answered.
5. Bundle yes/no questions after Q1 if the user is concise. A user who replies "y" / "n" / "skip" in 1-2 words tolerates a bundle. A user who replies with paragraphs wants the slow pace — keep one-at-a-time.

Every Q should earn its slot: removing it must leave the Requirements Document materially worse. If it wouldn't, cut the question.

Structured Elicitation Frameworks

Choose the framework that fits the requirement type. Use it to STRUCTURE the 5 Questions above, not replace them.

PICO Example (data feature):

P: Dashboard users monitoring real-time metrics
I: Add anomaly detection alerts
C: vs. current manual threshold setting
O: 30% faster incident detection (measurable KPI)

When to apply which:

• Feature Request → INVEST (ensures stories are sprint-ready)
• Data/Analytics/Research feature → PICO (forces measurable outcome definition)
• Product/UX feature → Jobs-to-be-Done (keeps focus on user motivation)
• Integration → 5 Questions only (frameworks add noise for plumbing tasks)

Step 2.5 — Ambiguity Scoring (Execution Gate)

After each question round, compute an Ambiguity Score to determine if requirements are clear enough to proceed. This prevents premature handoff to plan with vague inputs.

Scoring Formula

Ambiguity = 1 - weighted_average(dimensions)

Dimensions (weights vary by requirement type):
  Greenfield:  Goal (40%) + Constraints (30%) + Success Criteria (30%)
  Feature:     Goal (30%) + Constraints (30%) + Success Criteria (20%) + Integration (20%)
  Integration: Goal (20%) + Constraints (25%) + Success Criteria (20%) + API Contract (35%)

Dimension Scoring (0.0 – 1.0)

Threshold Gate

Scoring After Each Question

After each of the 5 Questions (Step 2), update the score:

Round 1 (WHO):    Goal ≈ 0.3, others = 0.0 → Ambiguity ≈ 91%
Round 2 (WHAT):   Goal ≈ 0.7, Success ≈ 0.3 → Ambiguity ≈ 72%
Round 3 (WHY):    Goal ≈ 0.9, Success ≈ 0.5 → Ambiguity ≈ 47%
Round 4 (BOUNDS): Constraints ≈ 0.6 → Ambiguity ≈ 30%
Round 5 (CONSTR): Constraints ≈ 0.9 → Ambiguity ≈ 12% ✅

If Ambiguity drops below 15% before all 5 questions are asked (e.g., user provides a detailed PRD), skip remaining questions and proceed. The gate is about clarity, not ceremony.

Display Format

After completing Step 2, show the user:

Clarity Score: [100 - ambiguity]%
  Goal:             [██████████] 0.9
  Constraints:      [████████░░] 0.8
  Success Criteria: [██████░░░░] 0.6  ← weakest
  Status: ACCEPTABLE (ambiguity 23%) — proceeding with noted gaps

Step 2.6 — CONTEXT.md Cross-Reference Gate

After elicitation, before hidden-requirement discovery, scan the user's answers for assertions about current behavior — phrasings like "the system X", "the code does X", "we already X", "right now it X".

For each such assertion:

1. Grep the codebase for evidence (function names, route handlers, schema definitions matching the asserted behavior).
2. Compare grep results to the user's claim.

This gate prevents the agent from silently transcribing user-asserted behavior that contradicts code — a common source of "the docs say X but the code does Y" drift.

Step 3 — Hidden Requirement Discovery

After the 5 questions, analyze for requirements the user DIDN'T mention:

Technical hidden requirements:

• Authentication/authorization needed?
• Rate limiting needed?
• Data persistence needed? (what DB, what schema)
• Error handling strategy?
• Offline/fallback behavior?
• Mobile responsiveness?
• Accessibility requirements?
• Internationalization?

Business hidden requirements:

• What happens on failure? (graceful degradation)
• What data needs to be tracked? (analytics events)
• Who else is affected? (other teams, other systems)
• What are the edge cases? (empty state, max limits, concurrent access)
• Regulatory/compliance needs? (GDPR, PCI, HIPAA)

Present discovered hidden requirements to user: "I found N additional requirements you may not have considered: [list]. Which are relevant?"

Step 3.5 — Completeness Scoring (Options & Alternatives)

When presenting options, alternatives, or scope decisions to the user, rate each with a Completeness score (X/10):

Always recommend the higher-completeness option unless the delta is truly expensive. With AI-assisted coding, the marginal cost of completeness is near-zero:

When showing effort estimates, always show both scales: (human: ~X / AI: ~Y). The compression ratio reframes "too expensive" into "15 minutes more."

Anti-pattern: "Choose B — it covers 90% of the value with less code." → If A is only 70 lines more, choose A. The last 10% is where production bugs hide.

Step 3.6 — Logic Consistency Check

After ambiguity + completeness pass, scan for cross-dimension contradictions. Ambiguity measures CLARITY of each dimension in isolation; this step measures CONSISTENCY across dimensions. A perfectly clear requirement can still contradict itself.

Checks

Run each, label verdict 🟢 pass / 🟡 warn / 🔴 fail:

Output Format

Logic Consistency Report:
  1. AC → User Story:      🟢 all AC trace to US-1 or US-2
  1b. EARS FR → story+AC:  🟢 FR-1..FR-5 each map to a story and an AC
  2. Business rule → AC:   🟡 "no duplicate emails" cited — exception flow missing
  3. Scope disjoint:       🟢
  4. Terminal states:      🟢
  5. Deps in constraints:  🔴 "PostgreSQL 15" missing from Q5 answer
  6. NFR measurable:       🟢
  7. Hidden reqs resolved: 🟢

Verdict: 1 🔴, 1 🟡, 5 🟢 → BLOCK handoff until 🔴 fixed

Gate Rule

Step 4 — Scope Definition

Based on all gathered information, produce:

In-Scope (explicitly included):

• [list of features/behaviors that WILL be built]

Out-of-Scope (explicitly excluded):

• [list of things we WON'T build — prevents scope creep]

Assumptions (things we're assuming without proof):

• [each assumption is a risk if wrong]

Dependencies (things that must exist before we can build):

• [APIs, services, libraries, access, existing code]

Step 4.5 — Functional Requirements (EARS)

Translate each in-scope item (Step 4) into atomic, testable functional requirements using EARS (Easy Approach to Requirements Syntax). EARS sits between user stories (WHY) and acceptance criteria (HOW we prove it) — it names exactly WHAT the system shall do, with an explicit trigger or condition.

Pick the simplest EARS template that fits each behavior; give each a stable ID (FR-1, FR-2, …):

FR-1  The API shall return responses in JSON.
FR-2  When a request omits a valid auth token, the API shall respond with HTTP 401.
FR-3  If the payment provider times out, then the system shall mark the order pending and queue a retry.

Rules (full guidance in references/ears-format.md):

• One shall per requirement — compound "shall do A and B" splits into two FR-n.
• Named subject, testable response — never "it shall handle gracefully". Move measurable limits into the response; pure performance targets stay in NFRs (Step 6).
• Don't smuggle HOW — observable behavior only; leave implementation to plan.
• Every FR-n traces up to a User Story and down to an Acceptance Criterion (checked in Step 3.6).

Skip EARS for Bug Fix and Refactor types (no new behavior to specify). For plumbing/integration, a few event-driven + unwanted-behavior lines suffice — don't pad with ubiquitous filler. EARS is a format recommendation, not a gate.

Step 5 — User Stories & Acceptance Criteria

For each in-scope feature, generate:

US-1 [P1]: As a [persona], I want to [action] so that [benefit]
  Independent Test: [specific action proving this story works end-to-end on its own —
                     e.g., "submit the form with valid data and see the saved record listed"]
  AC-1.1: GIVEN [context] WHEN [action] THEN [result]
  AC-1.2: GIVEN [error case] WHEN [action] THEN [error handling]
  AC-1.3: GIVEN [edge case] WHEN [action] THEN [graceful behavior]

Rules:

• Primary user story first, then edge cases
• Every story carries a priority: P1 = MVP-critical (the P1 set alone must be a viable, demoable product), P2 = important, P3 = nice-to-have. Implementing ONLY the P1 stories must still yield a working feature
• Every story carries an Independent Test — one concrete user action that proves the story end-to-end (through UI, logic, AND data if the story touches them). If no single action can prove it, the story is sliced wrong — split or merge until it can. This field is what downstream verification gates (plan's Coverage Gate, converge's gap scan, completion-gate's evidence trail) trace against
• Every user story has at least 2 acceptance criteria (happy path + error)
• Acceptance criteria are TESTABLE — they become test cases in Phase 3
• Each AC proves a specific FR-n from Step 4.5 — cite it (AC-1.2 → FR-3). Every unwanted-behavior FR (the If … lines) needs an error-path AC; that's where EARS earns its keep
• An AC whose THEN clause stops at the UI ("THEN the button shows a spinner") for a story that persists or fetches data is INCOMPLETE — the THEN must name the observable outcome ("THEN the order appears in the list / the record is persisted")

Step 5.5 — Key Entities (mandatory if feature involves data)

If ANY user story creates, reads, updates, or deletes data, list the key entities — WHAT the data is, not HOW it's stored (no table names, no column types, no ORM detail):

## Key Entities
- **Order**: what a customer submits — items, quantities, status (draft → submitted → fulfilled). One User has many Orders.
- **User**: the person placing orders — identity, contact. Referenced by Order.

Rules:

• One line per entity: name, what it represents, key attributes (plain language), relationships
• State-bearing entities name their lifecycle states — these become the state machine in requirements.mermaid
• This section SEEDS plan's data-model.md — a story that touches data but has no entity here is a spec gap (plan will bounce it back)
• Skip entirely for stateless features (pure computation, styling, config)

Step 6 — Non-Functional Requirements (NFRs)

Assess and document ONLY relevant NFRs:

Only include NFRs relevant to this specific task. Don't generate a generic checklist.

Step 6.5 — Tiered Recommendations

For product-oriented requirements (Feature Request, Integration, Greenfield), generate tiered strategic recommendations. This structures the path forward into actionable time horizons.

Three Tiers:

For each tier, specify:

### Quick Win (0-30 days)
- **Action**: [Specific deliverable]
- **Resources**: [Team size, tools, dependencies]
- **Expected Impact**: [Measurable outcome]
- **Risk if skipped**: [What happens without this]

### Differentiation (1-3 months)
- **Action**: [...]
- **Resources**: [...]
- **Expected Impact**: [...]
- **Risk if skipped**: [...]

### Long-term Moat (6-12 months)
- **Action**: [...]
- **Resources**: [...]
- **Expected Impact**: [...]
- **Risk if skipped**: [...]

Rules:

• Quick Win MUST be achievable in first sprint — no dependencies on later tiers
• Differentiation should create switching costs or unique capabilities
• Long-term Moat should compound (network effects, data moats, ecosystem lock-in)
• Every tier includes "Risk if skipped" — makes trade-offs explicit
• Skip this step for Bug Fix and Refactor types (no strategic dimension)

Step 7 — Artifact Triad

Produce three structured artifacts — not one prose doc. Plan consumes all three; each answers a different question.

Artifact 1: requirements.md

Structured document combining Steps 1-6.5. Save to .rune/features/<feature-name>/requirements.md:

# Requirements Document: [Feature Name]
Created: [date] | BA Session: [summary]

## Context
[Problem statement — 2-3 sentences]

## Stakeholders
- Primary user: [who]
- Affected systems: [what]

## User Stories
[from Step 5 — each with [P1|P2|P3] priority and Independent Test]

## Functional Requirements
[from Step 4.5 — EARS-format FR-n list; skip for Bug Fix/Refactor]

## Key Entities
[from Step 5.5 — mandatory if feature involves data; omit section for stateless features]

## Scope
### In Scope
### Out of Scope
### Assumptions

## Non-Functional Requirements
[from Step 6]

## Dependencies
[from Step 4]

## Risks
- [risk]: [mitigation]

## Strategic Recommendations
[from Step 6.5 — skip for Bug Fix/Refactor]

## Logic Consistency Report
[from Step 3.6 — verbatim, for audit trail]

## Next Step
→ Hand off to rune:plan (consumes all 3 artifacts)

Artifact 2: requirements.mermaid

Auto-generate from User Stories. Save to .rune/features/<feature-name>/requirements.mermaid.

Sequence diagram (primary happy path from US-1):

sequenceDiagram
  actor User
  participant System
  participant Database
  User->>System: [action from US-1]
  System->>Database: [read/write from AC-1.1]
  Database-->>System: [response]
  System-->>User: [result from AC-1.1]

State machine (only if any User Story implies state):

stateDiagram-v2
  [*] --> initial
  initial --> processing: [trigger from AC]
  processing --> success: [happy path AC]
  processing --> failed: [error AC]
  success --> [*]
  failed --> initial: retry

Skip state machine if feature is stateless (simple CRUD with no lifecycle). Sequence is always produced.

For a polished editorial visual, diagram can redraw requirements.mermaid (sequence → type sequence; state machine → type state) as self-contained HTML. Offer it when the reader would learn more from a designed figure than from raw Mermaid — suggested_next: diagram.

Artifact 3: tasks.md

Pre-broken implementation tasks grouped by user story (vertical slices), NOT by layer. Plan refines this backbone, does not create from scratch. Save to .rune/features/<feature-name>/tasks.md:

# Implementation Tasks: [Feature Name]

## Foundational (blocking prerequisites)
> Shared infrastructure NO story can start without. Keep minimal.
- [ ] Schema/migrations for Key Entities shared by 2+ stories
- [ ] Auth/routing/middleware skeleton (only if 2+ stories need it)

## US-1 [P1]: [story title]
> Layer order WITHIN the story: Data → Logic → Endpoint → UI → Test. UI never first.
- [ ] Data: [entity/migration this story owns, if not Foundational]
- [ ] Logic: [each Business Rule this story enforces → one task]
- [ ] Endpoint: [handler/service/API this story calls — skip only if story is pure-UI]
- [ ] UI: [component/surface — calls the Endpoint task above; a UI element and the
      endpoint it calls are ONE story's tasks, never split across stories]
- [ ] Test: [each AC of this story → one test; happy path + error]
- **Checkpoint**: US-1 independently functional — run its Independent Test

## US-2 [P2]: [story title]
- [ ] ... (same internal structure)

## NFR Verification
- [ ] [each NFR from Step 6 → one measurement task]

Derivation rules:

• 1 User Story → 1 ## US-n section containing ALL its layers (data through test) — a story section with only UI/Endpoint tasks and no Logic/Data/Test is a broken slice; a UI task with no Endpoint task above it (and no pure-UI justification) is a dead button in waiting
• 1 Business Rule → 1 Logic task + 1 test task, inside the story that enforces it
• 1 AC → ≥1 Test task inside its story (happy path + error)
• 1 NFR → 1 NFR Verification task
• Foundational holds ONLY what 2+ stories share — story-specific work stays in the story
• P1 sections first; completing all P1 sections + Foundational = viable MVP

Why story-grouped: layer-grouped backbones ("all Data → all Logic → all Interface") invite the executor to finish the Interface layer and stop — shipping UI with no wiring. Story-grouped means every completed section is a demoable end-to-end slice.

Handoff

Emit signal to plan with paths to all three artifacts. Plan MUST read all three before producing phase files — the triad is the contract.

Step 7.5 — Glossary Sharpen (CONTEXT.md update)

After the artifact triad is saved, append/update the project glossary CONTEXT.md with any domain terms that were sharpened during this session.

1. Determine glossary location:
   • If CONTEXT-MAP.md exists at root → multi-context; pick the right per-context CONTEXT.md
   • Else if root CONTEXT.md exists → use it
   • Else if any term needs recording → create root CONTEXT.md lazily
   • Else → skip silently (no-op when no terms emerged)
2. For each new term, add a row to the Language table (term, definition, aliases-to-avoid, status).
3. For each user-asserted relationship, add to the Relationships section.
4. For each ambiguity surfaced during elicitation, add to Flagged ambiguities.

Conflict gate — if a new term has ≥0.7 token overlap with an existing one, surface to user (merge / rename / keep distinct). NEVER silently re-define an existing term.

Format reference: references/context-md-format.md.

Output Format

Triad of artifacts under .rune/features/<feature-name>/:

Inside requirements.md the Decision Classification table MUST appear verbatim — plan gates on Decision compliance, Discretion items skip approval:

Constraints

1. MUST ask up to 5 probing questions before producing requirements — never more, skip any you can infer from context
2. MUST prefer yes/no or multiple-choice questions — open-ended only when no reasonable option set exists
3. MUST NOT ask for information already present in the user's message, the repo, or classification — read/grep first, ask second
4. MUST cache each Q→A pair in the Requirements Document and reuse on subsequent BA sessions for the same feature
5. MUST identify hidden requirements — the obvious ones are never the full picture
6. MUST define out-of-scope explicitly — prevents scope creep
7. MUST produce testable acceptance criteria — they become test cases
8. MUST NOT write code or plan implementation — BA produces WHAT, plan produces HOW
9. MUST ask ONE question at a time by default; bundle yes/no batches only after user shows concise replies
10. MUST NOT skip BA for non-trivial tasks — "just build it" gets redirected to Question 1
11. MUST assign [P1|P2|P3] priority to every user story — the P1 set alone must be a viable, demoable product
12. MUST give every user story an Independent Test — one concrete action proving it end-to-end
13. MUST list Key Entities when any story touches data (Step 5.5) — a data-touching story with no entity is a spec gap
14. MUST group tasks.md by user story (vertical slices), never by layer

Returns

Sharp Edges

Known failure modes for this skill. Check these before declaring done.

Done When

• Requirement type classified (feature/refactor/integration/greenfield)
• 5 probing questions asked and answered (or extracted from spec/PRD)
• Ambiguity Score computed and displayed — must be ≤ 40% before proceeding (≤ 25% preferred)
• Hidden requirements discovered and confirmed with user
• Scope defined (in/out/assumptions/dependencies)
• Functional requirements written in EARS format (FR-n) — skip for Bug Fix/Refactor
• User stories with testable acceptance criteria produced, each AC citing the FR-n it proves
• Every story has a [P1|P2|P3] priority and an Independent Test; P1 set alone = viable MVP
• Key Entities listed (if feature involves data) — seeds plan's data-model.md
• tasks.md grouped by user story (vertical slices), Foundational section minimal
• Non-functional requirements assessed (relevant ones only)
• Logic Consistency Report produced — 0 🔴 before handoff (🟡 logged as Risks)
• Tiered recommendations generated (Quick Win / Differentiation / Moat) — skip for Bug Fix/Refactor
• Artifact triad saved: requirements.md + requirements.mermaid + tasks.md
• Out-of-scope match check completed (verdict logged: no-match | similar | exact-match-overridden | exact-match-accepted)
• Handed off to plan for implementation planning

Cost Profile

~3000-6000 tokens input, ~1500-3000 tokens output. Opus for deep requirement analysis — understanding WHAT to build is the most expensive mistake to get wrong.