Ground

<!-- mirror:transform — intentional divergence: private prediction-infra names genericized (the private copy wires a specific ledger + scheduled grader) -->

The law this skill runs on: a product dies not from bad craft but from being aimed at no one. Your whole suite optimizes SUPPLY — /frontier pushes craft to the 2026 standard, /harmonize enforces consistency, /raze and /moonwalk perfect the code, /taste encodes your aesthetic — and every one of them makes the product more impressive. None checks whether anyone wants it. Run as designed, the suite manufactures the exact failure you fear: a frontier-grade thing the market is indifferent to ("generation is free; judgment isn't" — and the deepest judgment is what to build). ground is the missing demand anchor. It forces each bet's unstated demand assumption explicit, classifies the evidence against a hard hierarchy that refuses to let attention masquerade as demand, and — in the common case — hands you the cheapest test that would settle it rather than a greenlight. It makes "someone wants this" a falsifiable, graded belief instead of a founder's conviction.

Three hard invariants:

• PULL is earned, never assumed — cite the demand or declare UNKNOWN. Evidence is ranked: Tier 0 (revealed & costly — someone paid, switched, retained past week 1; a competitor's real paying/churn numbers; usage on your own thing), Tier 1 (specific, recent, unprompted stated pain a named person would pay to fix), Tier 2 (vanity — search volume, funding, hype, "market size", upvotes, competitor marketing). A PULL verdict requires ≥ Tier-1 evidence, cited. Vanity proxies establish attention, never demand — they can only DOWNGRADE; alone they cap the verdict at UNKNOWN. The default is UNKNOWN/PUSH, and the common deliverable is the cheapest test that would upgrade it — not a verdict.
• Falsification framing + founder-bias guard. The research runs as the BEAR case — argue it flops, find the graveyard of similar dead products, find who does NOT want it; the proposer is never the judge. ground names the feature YOU find most impressive precisely to flag it: supply-pull is not demand-pull, and the more impressive you find it, the higher the PUSH suspicion. PULL is earned by surviving disconfirmation, never by a fluent market narrative.
• Gates and grades — never just a document. The verdict is a required input to the ship/build decision: a build of a PUSH or UNKNOWN bet without a logged demand test is flagged (enforce the gate via /redline). Each bet carries a falsifiable demand-prediction: a grader-measurable one is emitted to your prediction ledger (auto-flags on a FALSE grade); a MANUAL-measured one (retention, install→2nd-run, acted-on-push rate — which an automated grader cannot read) is recorded in the playbook with a manual-grade date. Never emit a metric the grader cannot read (the placement-split trap). The playbook is outcome-bound, not a passive doc nobody reads at decision time (the false-CLEAN trap).

/ground <bet|product> (default — extract + falsify + verdict)

Fan the research to a scout subagent; only the cited evidence + verdict return to main context.

1. Extract the implicit demand bet — and decompose a portfolio. Restate what is being built/planned AND the unstated assumption underneath it: "we believe the market wants Y / will pay / switch / adopt because Z." If the target is a PRODUCT or a bundle (not a single feature), DECOMPOSE it into its constituent bets and ground each SEPARATELY — never average a bundle into one verdict; the bundle-as-one-bet framing is the weakest and hides which parts are PULL inside a PUSH (and vice versa). Ground the most supply-loved / riskiest bet first. Name the feature you find most impressive — the founder-bias flag (invariant 2).
2. Falsify it (the bear case). Run /deep-research as the executor, framed to DISPROVE demand: the dead similar products and why they died, who does not want this, the reasons it stays a demo. Build the strongest disconfirmation, not a balanced survey.
3. Classify the evidence by tier, citing each piece. Tier 0 / Tier 1 / Tier 2. Vanity (Tier 2) can only downgrade. No citation = it does not count.
4. Render the verdict — exactly one, with its evidence:
   • PULL — ≥ Tier-1 demand evidence, cited. The only verdict that licenses a build now.
   • PUSH — supply-attractive but demand-unsupported (you looked with a falsification frame and found no Tier-0/1 support). The danger profile — flag loudly; not a proof of doom, a reason to test.
   • UNKNOWN — only vanity / insufficient signal.
   • DRIFTED — prior support no longer present in recent Tier-0 behavior (demand moved since the playbook).
5. When sub-PULL, name the settling signal, then design the cheapest test (the common deliverable). First name the SINGLE revealed-behavior micro-signal that would flip the verdict — the one number that settles it (e.g. install→ran-2nd-command, free→paid, week-4 acted-on-push retention, your own open-and-act rate). Then from the fixed cheap-test menu — landing/smoke test, pre-sale or LOI, fake-door, 5 specific-pain interviews, concierge MVP — pick the cheapest experiment that produces THAT signal. This is the output in the common case, NOT a greenlight.
6. Propose the playbook delta + the prediction. Write nothing yet: surface the verdict, the tier-cited evidence, the test (or, for PULL, the bet to commit) and the falsifiable demand-prediction it would emit. If a SHIP is in play, route the gate to /redline.

/ground audit

Read-only. Re-grade the existing playbook against the LATEST signals: surface PUSH bets currently being built or planned (impressive-but-unwanted), DRIFTED bets (demand moved), and demand-predictions now due for grading. Each with its tier-cited evidence or the explicit "only vanity / no signal." Propose and write nothing. The "is our roadmap still aimed at demand, and has the market moved under us" check — mirrors /raze, /hone, /harmonize audit.

/ground apply

Execute ONLY approved updates from a prior scan. Commit the playbook delta — the demand thesis, dated, each bet tagged PULL/PUSH/UNKNOWN/DRIFTED with its tier-cited evidence — and emit GRADER-MEASURABLE demand-predictions into your prediction ledger (ride an existing scheduled grader; a new metric kind should need zero new scheduler wiring); a MANUAL-measured metric (retention / install→2nd-run / acted-on-push) is recorded in the playbook with a manual-grade date instead, never handed to a grader that would never score it. For a PUSH/UNKNOWN bet being built anyway, install the demand-test gate via /redline. One bet = one commit. Record the running metric (PUSH-caught, prediction hit-rate) in the playbook's ledger.

Rules

• Evidence hierarchy is law: PULL ≥ Tier-1, vanity caps at UNKNOWN, cite or declare UNKNOWN — never infer demand from attention. Search / funding / hype establish that people are LOOKING, never that they will PAY, SWITCH, or STAY.
• The default is UNKNOWN, and UNKNOWN ships a test, not a greenlight: generate the signal with the cheapest experiment; do not read the vanity and call it demand. For a pre-PMF project this is the point, not a limitation — conviction becomes evidence.
• A sub-PULL verdict ends the craft lane in-session: once PUSH/UNKNOWN/DRIFTED is rendered, the only next artifact for that bet is the named cheapest test (or its result) — no further drafts, polish, essays, repoints, or analyses "to be sure." Producing more craft after the verdict is the gate violation happening live: name it and stop. Needing ever-more analysis to keep the bet standing is the kill signal, not an analysis gap.
• Decompose portfolios; never average: a product is a set of bets — ground each separately and report per-bet verdicts; a single averaged verdict on a bundle hides the PULL inside a PUSH (and the PUSH inside a PULL). Always name the settling micro-signal before designing a test.
• Falsification + founder-bias guard: the research builds the bear case; the proposer is not the judge; the more impressive you find it, the higher the PUSH suspicion; PULL is earned by surviving disconfirmation.
• Gates and grades, not a passive doc: the verdict is a required input to ship (a build of a PUSH/UNKNOWN bet without a logged demand test is flagged — enforce via /redline); each bet carries a falsifiable demand-prediction — emitted to the prediction ledger if grader-measurable (auto-flag on FALSE), else recorded in the playbook for manual grading.
• Recency through the same hierarchy: DRIFTED needs Tier-0 revealed movement; a recent Tier-2 spike is noise → UNKNOWN, not DRIFTED. No special pass for "but it's recent."
• Lane discipline: ground is the DEMAND axis; /frontier is the SUPPLY/craft axis (is it good? — they pipeline: frontier makes it excellent, ground aims the excellence at demand); /deep-research is ground's research executor (do not re-roll it); /redline enforces the ship gate; your prediction ledger grades the demand bets; /taste is YOUR internal aesthetic (a supply preference), never the market's demand. Your ship/no-ship decision consumes ground's verdict. A project fact → /burn rent.
• Propose-on-approval, reversible: default writes nothing; playbook, prediction, and gate land only in apply, one bet per commit, dated and tier-cited.
• Essentialness metric (honesty hook): PUSH bets caught before build (impressive-unwanted features killed/redirected pre-investment) + demand-prediction hit-rate. Never flags a PUSH and never beats a coin flip on predictions = astrology → /raze it. It earns its keep the first time it stops you sinking a month into a frontier-grade feature the market was never reaching for.

Pipeline placement: /frontier (make the craft excellent) → ground (prove the excellence is aimed at real demand, or hand back the cheapest test) → build → prediction grading (grade the demand bet against reality) → re-ground on drift. ground is the demand anchor on the whole flywheel; without it the suite's supply-side power compounds in a direction nobody asked for. Alias: /pmf runs this skill — ground is the self-documenting verb (stay grounded in evidence, do not float on proxy and conviction); pmf is retained as the recognizable term.