Validate Idea

Harness adherence (non-negotiable)

• Follow the step sequence for the selected mode exactly. Do NOT skip steps or reorder them.
• Each step file is a contract. Read it fully before executing. Step files override general guidelines.
• Mode detection is automatic. The router decides create/edit/validate. Do NOT ask the user which mode unless the intent is genuinely ambiguous.

Mode selection

Explicit user intent always wins over file presence. If the user says "stress-test this fresh idea" while a stale brief sits on disk, that is create mode — say so and proceed. File presence is a default, not a veto.

Only validation-brief.md is a mode sentinel. idea-brief.md is an entry input, never a sentinel: it is written by create mode's skip-to-exploration path and by explore-idea, so keying mode detection on it makes create mode unreachable after the first run — a one-way lock into edit/validate. Do not repeat that mistake.

Problem-first entry: artifacts/output/02-research/problem-space-brief.md from /unpack-problem is accepted as the idea to validate. It routes to create mode with the problem brief's selected solution concept as the framed idea (see steps/step-03a-idea-framing.md) — it never routes to edit or validate, even if a validation-brief.md also exists. The problem brief is the newer artifact and wins; say which one you adopted.

If intent is genuinely ambiguous, ask: "Are you starting a new idea, refining an existing brief, or stress-testing it?"

Mode routing + steps

Step 0 — Scope Gate (non-skippable): anchor scope first via steps/step-00-scope-and-decision-anchoring.md; the selected mode is the locked track. step_tracker.js exits 1 if Step 1 begins without a locked track. Intent triage at gate: Clear (C ≥ 0.85) runs the locked track; Ambiguous (0.50–0.85) halts to a 2–3 Track Fork; Trivial (C < 0.50) executes directly. Gate enforced in all step_tracking modes including off (DNA 4).

• Create (7 steps) → steps/step-01a-session-setup.md → steps/step-02a-input-analysis.md → steps/step-03a-idea-framing.md → steps/step-04a-stress-test-r1.md → steps/step-05a-stress-test-r2.md → steps/step-06a-go-pivot-kill.md → steps/step-07a-handoff.md
• Edit (5 steps) → steps/step-01b-load-existing.md → steps/step-02b-identify-gaps.md → steps/step-03b-revise.md → steps/step-04b-stress-test.md → steps/step-05b-finalize.md
• Validate (5 steps) → steps/step-01c-open-questions.md → steps/step-02c-7-branches.md → steps/step-03c-cross-branch-check.md → steps/step-04c-decision-log.md → steps/step-05c-lock-handoff.md

When to invoke

• The user has a concept, pitch, or feature idea that nobody has stress-tested, and wants it challenged before research or engineering time is committed
• The user says "stress-test this idea", "poke holes in this", "validate this concept", or invokes /validate-idea
• A brief already exists and the user wants it refined (edit mode) or interrogated question-by-question (validate mode)
• A problem-space-brief.md from /unpack-problem needs its selected solution concept validated (problem-first entry, create mode)

When NOT to invoke

• The user has a problem, not an idea — use /unpack-problem; there is nothing to validate yet
• The concept has already passed validation and the ask is market evidence — use /explore-idea
• The ask is to structure raw notes into a brief — use /shape-up
• The ask is to generate concepts, not test one — use /brainstorming
• The ask is to interrogate one claim (this premise, this architecture bet, this estimate) rather than judge a whole concept — use /grill-me. A panel is cheaper than a wrong GO.
• The ask is how to run a validation experiment (smoke test, concierge MVP, landing page) after the concept is already validated — use /validation-patterns
• Multi-stakeholder or irreversible concepts where a solo Socratic pass is not enough seats — the 7-branch validate mode is a single-agent review; escalate to /round-table for genuine opposition between personas

Prerequisites

• Create mode: none (this is the entry point)
• Edit mode: artifacts/output/01-discovery/validation-brief.md exists
• Validate mode: artifacts/output/01-discovery/validation-brief.md exists
• idea-brief.md is an accepted entry input in edit/validate mode when no validation-brief.md exists yet (read-only — never write to it; see Mode selection on why it is not a sentinel)
• Problem-first entry: artifacts/output/02-research/problem-space-brief.md from /unpack-problem is accepted as the idea to validate — routes to create mode

Halt conditions

• [NO-GO] verdict issued (create mode) — the autopsy and a route back to problem discovery are the whole output; no build steps
• Scope-dispute arbitration: PM vs. tech-lead scope conflicts escalate to @founder per .agents/workflow.md §3.2 Escalation Ladder Level 3 (binding).
• Cross-branch contradiction unresolved (validate mode)
• Revision-round limits without convergence (edit mode): per-section rounds per steps/step-03b-revise.md, stress-test resolution rounds per steps/step-04b-stress-test.md

Anti-sycophancy gate (SPC)

If the user's answers across a stress-test round are uniformly positive — every push lands a confident answer, no red flags surface, no assumption cracks — do not proceed to the next round or verdict on momentum. Inject one adversarial question built from the strongest counter-hypothesis the answers imply (e.g., "Everything you've said says this should already exist — why doesn't it?"). If the answer survives, record it as validated evidence; if it cracks, that crack is the finding. A stress-test with zero friction has tested nothing. The step files apply this gate at their checkpoints (step-04a, step-04b, step-05a).

Emit the marker. Every round that survives the gate records one machine-readable line, so a zero-friction round is visible in the transcript instead of hiding in prose:

[SPC: <one-line adversarial probe> → <survived | cracked: what gave>]

A round that closes with no [SPC:] line and no red flag is an unexamined round. Coverage is exercised by evals/suites/invariants/validate-idea-no-go.json.

Session durability (long Socratic sessions)

Create mode can span 7 rounds and dozens of answers. Nothing is written to disk until step-07a, so a crash, a context truncation, or an abandoned session loses every answer. Checkpoint as you go:

• After each round (end of step-04a, step-05a), append the answers gathered so far to artifacts/output/01-discovery/validation-brief.md under a ## Diagnostic log (in progress) heading. Say plainly that the brief is incomplete.
• On abandon or interruption — user says stop, goes quiet, or the session ends without a verdict — the in-progress brief plus the Q1–Q7 coverage table (unasked questions dispositioned DEFERRED + owner + revisit trigger) is the deliverable. Persist it and tell the user where it is. A partial brief with honest coverage beats a complete brief invented from summary.
• Never issue a verdict from a truncated answer set. Missing evidence makes a criterion PARTIAL, which forces [RESHAPE] (see steps/step-06a-go-pivot-kill.md).

State machine integration

At start: run node .agents/scripts/orchestrator_state.js status At end: run node .agents/scripts/orchestrator_state.js complete --agent founder --artifact 01-discovery/validation-brief.md

Memory integration

At start: Load founder context before any Socratic analysis:

@memory-controller load founder [validate-idea — {idea name}]

At completion: Write verdict and rationale to memory, then session summary — mandatory:

@memory-controller write active-decisions.md
### [PRODUCT] Idea validation verdict — {idea name} [date: YYYY-MM-DD] [agent: @founder]
Verdict: [GO] When: <NOW|GATED|NEXT-CYCLE> | [RESHAPE] | [NO-GO]
Survived: everything | the need, not the mechanism | nothing
Rationale: {1-2 sentences on why, in plain language}
Key assumptions validated: {list}
**Status:** active

@memory-controller session-write [agent: @founder]
Worked on: Idea validation — {idea name}
Decisions: {plain sentence: build this / keep the why replace the how / don't build this} — [GO|RESHAPE|NO-GO]. {key rationale}
Next step: {explore-idea if [GO], re-run validate-idea if [RESHAPE], unpack-problem if [NO-GO]}
Blockers: {unresolved assumptions or "none"}

Verdict tokens stay bracketed in memory too — a bare Verdict: NO-GO is unparseable downstream.

Output artifacts

Full paths, so step_tracker.js audit can decide whether this skill ran untracked:

• artifacts/output/01-discovery/validation-brief.md — create mode (canonical)
• artifacts/output/01-discovery/shaped-brief.md — when entered from /shape-up
• artifacts/output/01-discovery/idea-brief.md — brownfield/edit mode only; never a create-mode output (writing it would poison mode detection on the next run)

This section exists because audit --all cross-checks declared artifact paths against step-audit.json. Without it, a skill with zero step records and artifacts on disk is indistinguishable from one that never ran — and this skill was measured at 18 step files and zero records, which nothing reported.

Done when

• Brief written to artifacts/output/01-discovery/validation-brief.md following .agents/templates/discovery/validation-brief-template.md (canonical output — see steps/step-07a-handoff.md). idea-brief.md is never a create-mode output — writing it would poison mode detection on the next run.
• Every Q1–Q7 question carries a terminal state (COVERED / DEFERRED / REJECTED / OUT-OF-SCOPE) — no blanks
• The brief (the document surface) ends with the plain sentence and one bracketed verdict card at the depth the stakes require; Survived:, plus When: on [GO] — format in .agents/references/vespyr-dna.md §Decision Gate
• The chat reply carried only the plain sentence, the consequence, and the next action — no token, no card field, no marker (.agents/references/vespyr-dna.md §Two surfaces)
• [NO-GO] → No-Go Autopsy emitted, alternatives omitted, next action routes to problem discovery
• Handoff to next phase is unambiguous