Card proposal: think-verify (empirical reality-check move)

[franck-ramarojohn]

Which instruction I repeat

Across four unrelated projects (a construction time-tracking SaaS, a rental-management CRM, an internal RBAC platform, a webinar funnel) my single most repeated directive to the agent is a variant of: "before you tell me it works, or before we act, go check the live system and tell me where reality differs from the claim."

Examples, paraphrased from real sessions:

• "Are you 100% sure this won't break anything, can I upload and replace?"
• "The email promises coaching, but we only offer written support." (claim vs implementation)
• "Verify on HubSpot / n8n before telling me it works."
• "Test this index on staging to settle it."

grill challenges my reasoning. Nothing in the deck confronts a claim with the live system. That gap is what this card fills.

Proposed card

Name: think-verify
Type: move

Use when: A claim, promise, or assumed state must be true before the user relies on it or acts.
Works on by default: The current claim, promise, message, or assumed state.
Effect: Read the real state from available live sources (code, logs, database, queries, UX, external platform), compare it to the claim, and enumerate the divergences.
Result: A grounded verdict, conforming or a list of gaps, backed by observed evidence.
Duration: One agent turn.
Limits: Separate observed fact, inference, and unverifiable claim. Do not conclude from reasoning alone when a live source is reachable. Do not decide for the user.
Combines with: Works on a selected focus or the preceding move's result. Pass the verdict to another move, an output, or modifiers.

Flow

flowchart LR
    A["Claim or assumed state"] --> B["Identify reachable live sources"]
    B --> C["Read the real state"]
    C --> D["Compare claim to reality"]
    D --> E["Enumerate gaps"]
    E --> F["Return grounded verdict"]

Loading

If no live source is reachable, say so explicitly and mark the claim as unverified.

Format

Begin the combo trace with > 🎯 **<focus>** → 🔬 **VERIFY**, followed by Verified, Gaps, Unverifiable, Verdict.

Why this is not grill

• grill is multi-turn and adversarial against a decision's reasoning; think-verify is one-shot and empirical against observable state.
• grill asks the user demanding questions; think-verify reads sources and returns a gap list, asking nothing.
• They compose: verify establishes what is true, then grill pressure-tests the decision built on it.

Possible overlap to discuss

If maintainers prefer, the "confront a promise or message against reality" facet could stay here, while a separate reconciliation facet (diff current state against an authoritative reference, such as a prior system or spec) becomes its own think-reconcile. Happy to split.