Discover

Discover — Conversational

<!-- BUDGET_EXCEPTION: Entry-point conversational skill documents all three depth tiers (Light / Medium / Deep) in one body. Compacting below ~2,500 tokens removes structural meaning. Auto-downgrade routes simple invocations to Light; only Medium/Deep depth loads the procedural references. -->

Transform vague ideas into shared clarity through adaptive conversation. Capability metadata + load map: routing.yaml. Methodology (conversation-IS-the-alignment, blunt-peer stance, equal-weight rule, resolution-exit-not-patience-exit) + when NOT to use: references/playbook.md.

Core philosophy: Just talk with your agent. Close the gap between stated requirements and true needs through conversation — not documents, formal phases, or plan mode.

Core question: What would we silently get wrong if we just started building?

How It Works

You describe what you want → agent scans context + assesses complexity silently → adaptive questions begin → conversation continues to mutual clarity → build directly or save a spec. No plan mode. No pipeline stages. No mandatory artifacts. The conversation IS the alignment.

Before Starting

Apply references/_shared/before-starting-check.md. Then:

• Mode resolution (references/_shared/mode-resolver.md). budget: fast; Adaptive Depth auto-calibrates Light / Medium / Deep. Overrides: "quick scope" / "deep interview" / "just ask 3 questions". --fast collapses to Light depth — skips operator-craft stance load, skips idea-critic gate, skips 5 mandatory spec sections on save (light_spec: true).
• Run Step 1 Context Gathering per references/procedures/context-gathering.md — scan codebase + docs/forsvn/artifacts/ + experience docs + learned rules + out-of-scope + project CLAUDE.md. Anything found = question you don't ask.
• Load operator-craft stance (3 playbooks: ceo-cognitive-patterns + yc-six-forcing-questions + minimalist-entrepreneur) for non-trivial work; match founder-domain frames per the matrix in context-gathering.md.

Adaptive Depth

Signal	Depth	Behavior
Clear task, existing codebase, well-defined scope	Light (3-5 questions)	Surface assumptions, lock scope, go
Feature with some ambiguity, multiple approaches	Medium (5-10 questions)	Explore key decisions, probe edge cases
Vague idea, greenfield, "I want to build X"	Deep (multi-round)	Challenge premise, interview across zones, iterate until the Step 6 concreteness gate passes

Depth sets the question ceiling; the Step 6 concreteness gate sets the exit. Deep is multi-round until the gate passes — not question-count-bounded. Light/Medium ceilings stand: at ceiling with open dimensions, ask per-dimension sign-off, wrap done_with_concerns. A fully-concrete brief passes the gate on round one — no extra grilling. "That's enough, let's build" triggers the sign-off ask, not a silent skip.

Execution

Per references/procedures/orchestration-steps.md (full step-by-step including Premise Check, Mode Detection, Idea Critic Gate, Adaptive Coverage Zones, Conversation, Clarity Check, Output):

• Step 1 — Context Gathering (silent) per procedures/context-gathering.md.
• Step 2 — Premise Check (non-trivial work): restate outcome, "what if we did nothing?", what already exists. Framing checkpoint after first answer. Skip when scoped / continuing / obviously sound.
• Step 2.5 — Mode Detection (Idea Stage vs Plan Review) one-shot read; on ambiguity, ask one question. Plan-Review uses 4-Mode Framework picked once via AskUserQuestion and locks Step 7 verdict vocabulary — see procedures/plan-review-modes.md.
• Step 2.6 — Divergence Pass (idea-stage, conditional). Generative ideation before convergence — fires only on Deep depth + open solution space (or operator "brainstorm this"); never on Light or plan-review; offered not forced when a committed solution was brought. Generates 5-8 candidate directions across explicit axes (judgment suspended), converges to a recommended shortlist, then rejoins Step 2.7. Distinct from Premise Check (reframe), idea-critic (evaluate), debate-agents (pressure-test known options — divergence feeds it). Full: procedures/divergence.md.
• Step 2.7 — Idea Critic Gate (idea-stage only). One dispatch of agents/idea-critic.md scored against 5 red + 5 green flags. Threshold: ≥2 red OR <2 green → PUSH_BACK. Full contract: procedures/idea-critic-dispatch.md.
• Step 3 — Adaptive Coverage Zones (3-5 zones for THIS problem, not 5 fixed dimensions). State zones upfront. Zone library by domain (product feature, business strategy, marketing initiative, infra/devops, design task) in procedures/orchestration-steps.md.
• Step 4 — Conversation: interview techniques per procedures/interview-techniques.md; communication discipline (banned phrases, take-a-position rule, always-recommend rule, pushback patterns) per procedures/communication-discipline.md; domain probing via lazy references/question-bank.md loads.
• Step 5 — Complex Decision Points → invoke debate-agents as sub-routine when 2+ viable approaches with non-obvious tradeoffs + decision expensive to reverse. Not when clear best answer / strong preference / easily reversible.
• Step 6 — Clarity Check + Concreteness gate: six core requirement dimensions (problem, audience, constraints, success criteria, scope, anti-goals) each marked concrete or open; wrap is blocked while any is open unless the operator signs off on that specific dimension — sign-off is per-dimension, never blanket; no downstream handoff while the gate fails. Then summarize decisions, playbook-citation self-check, verdict assignment (idea-stage: VALIDATED / NEEDS_MORE_VALIDATION / PIVOT; plan-review: one of 7 verdicts mapped to chosen mode). Resolution-exit (replaces patience-exit) per procedures/orchestration-steps.md § "Resolution-exit" — 3 conditions, with N/A clauses for Light / Contract / HOLD SCOPE / Premise-skipped sessions. Operator override allowed; silent exit is not — signed-off dimensions + unresolved branches log inline + persist (## Concreteness Checklist, ## Open Branches (operator-overridden)) with status: done_with_concerns.
• Step 7 — Output. Default: conversation context. Save only when explicitly asked / session ending / external consumer / natural milestone + confirmation. Formats: operator-grade spec (5 mandatory sections), Light-depth compact, Contract format, Handoff plan (path-free, when a fresh agent/session is the next consumer). Full: procedures/output-formats.md.

Artifact Contract

• Path: docs/forsvn/artifacts/meta-discover-<YYYY-MM-DD>-<slug>.md (flat v2; working draft, only on user save).
• Lifecycle: spec — iterated until promoted to task-breakdown or system-architecture.
• Frontmatter: skill, version, date, status, stack (=meta), review_surface (=md), decision_state, review_tool, reviewed_at, reviewer, mode, plan-review-mode, light_spec. Full template: procedures/output-formats.md.
• Required sections (Medium/Deep): Premise Challenge · Dream State Mapping · Decided Approach · Implementation Alternatives (min 2-3) · Temporal Interrogation · Key Decisions · Edge Cases · Failure Conditions · Out of Scope · Open Questions · Concreteness Checklist (both formats — Step 6 gate state) · Open Branches (only if done_with_concerns) · Implementation Notes · Verdict. Light-depth saves skip 5 sections (Premise Challenge, Dream State Mapping, Implementation Alternatives, Temporal Interrogation, Verdict) — compact format light_spec: true.
• Consumed by: operator (audit trail) · breakdown-tasks · architect-system · review-work (scope-drift detection).
• Side effect (idea-stage only): spawns agents/idea-critic.md once at Step 2.7.

Context Resolution Order

When discover (or any downstream skill) needs prior decisions: (1) conversation context, (2) artifact on disk, (3) discovery — ask or scan. Downstream skills don't require files; they need decisions known, from any source.

Configuration

Full configuration table (depth, mode, plan-review-mode, output, zones, idea-critic): references/configuration.md.

Anti-Patterns + Edge Cases

references/anti-patterns.md — 10 interview + 7 resolution/exit + 4 output + 12 edge cases. Read before any answer-handling moment that smells off.

Skill Deference

• FEATURE / TASK to clarify → this skill.
• Declining METRIC → diagnose.
• Multi-perspective debate → debate-agents (or Step 5 sub-routine).
• Know what to build, need design → architect-system.
• Decompose into tasks → breakdown-tasks.

Chain Position

Previous: none (or any skill that surfaces a need for clarification). Next: architect-system, breakdown-tasks, or direct implementation. Re-run: requirements change significantly, new constraints, or implementation reveals the spec was wrong.

Durable Rules (protected)

<!-- SLOW_UPDATE_START --> <!-- No pinned rules yet. Populate via the slow-update workflow (see references/slow-update-fence.md). Each pinned rule must (a) be procedural not instance-specific, (b) be earned from a regression or critic-flagged failure, (c) cite the artifact / decision record that justified pinning. --> <!-- SLOW_UPDATE_END -->

Completion Status

• DONE — discovery converged, decision clear (optionally saved).
• DONE_WITH_CONCERNS — decision made but non-blocking open questions or explicit caveats; flagged inline (and pinned to spec frontmatter if saved). Also fires when operator overrides resolution-exit (Open Branches section non-empty) or signs off an open dimension (Concreteness Checklist shows open — signed off).
• BLOCKED — irreconcilable conflict in inputs or scope; needs human resolution.
• NEEDS_CONTEXT — user cannot answer key questions; recommend upstream skill (research-icp, research-market, diagnose).

References

• references/playbook.md — why, methodology, when NOT to use
• _shared/ — before-starting-check, mode-resolver
• procedures/ — context-gathering, communication-discipline, interview-techniques, divergence (Step 2.6 generative pass), idea-critic-dispatch, output-formats, plan-review-modes, orchestration-steps (full Step 2-7 detail), resolution-exit conditions
• references/anti-patterns.md, question-bank.md, example-contracts.md, configuration.md
• references/operator-playbooks/ — 9 practitioner-grade frames (3 always-on stance + 6 founder-domain on product-context match)
• agents/idea-critic.md — sub-agent dispatched at Step 2.7 idea-stage