agent-grilling

name	agent-grilling
description	Use only when an active workflow routes here and a goal, requirement, design direction, plan boundary, or lane split may hide omitted actors, flows, edge cases, assumptions, dependencies, contradictions, risks, or second-order effects before downstream work. Do not use for review or returned-lane integration.

Agent Grilling

Discover what the current formulation has not noticed. Expand the problem model, trace high-impact gaps to their consequences, and return formulation or requirement deltas. Grilling and challenge are core probes in this deep investigation; the outcome is discovery, not a stress-test verdict or artifact review.

Iron Law

NO HANDOFF WHILE A HIGH-IMPACT BLIND SPOT REMAINS UNSEARCHED OR UNDISPOSITIONED.

Resolve source- and agent-answerable gaps before asking the user. Escalate only true decisions such as product direction, taste, priority, privacy/cost, destructive/public action, account access, or unavailable user-owned facts.

Use For

• clear-looking or fuzzy formulations with unexamined behavior or consequences
• requirements that need enrichment before Spec
• likely omitted actors, flows, states, permissions, data rules, or failure paths
• user feedback like "not usable", "hard to scan", or "feels wrong"
• cross-cutting changes, migrations, integrations, or risky lane boundaries

Use agent-debate to compare contested alternatives. Use agent-review to judge an existing artifact. Use agent-lanes only after boundaries are discoverably clear.

When using agents for the pass, use agent-runtime for capability, authorization, and session lifecycle.

Method

model the stated intent and current assumptions
  -> inspect source context that can constrain it
  -> sweep the affected dimensions for what is missing
  -> turn omissions into explicit gap questions
  -> grill high-impact gaps and proposed answers to causes and consequences
  -> resolve them from source, focused agents, or safe assumptions
  -> write the formulation and requirement deltas
  -> repeat for newly exposed high-impact gaps
  -> return deltas to the active calling owner

Sweep only affected dimensions, not a ceremonial checklist:

actors and permissions | end-to-end flow and state transitions
data ownership and lifecycle | edge, failure, recovery, and abuse paths
compatibility and migration | dependencies and operations
success evidence | boundaries and non-goals

Treat the first plausible answer as the start of grilling, not the result. For every high-impact candidate, establish:

gap -> proposed answer -> why -> evidence -> underlying assumption or cause
    -> counterexample -> boundary -> failure or second-order effect
    -> resolver -> disposition -> requirement or decision delta

Allowed dispositions are resolved from source, safe assumption, draft requirement, constraint/risk/non-goal, true user decision, or deferred with reason and impact. Do not stop at a question list.

Example: “invite teammates” hides whether existing private notes become visible. Tracing ownership and migration exposes a privacy failure; the requirement delta is that migration must not broaden visibility and sharing needs an explicit move or product decision.

Output

Return the discovery delta, not a polished restatement:

Discovery snapshot handle / same-scope target:
Enriched formulation:
Discovered gaps and dispositions:
New or changed requirements / decisions:
Candidate additions / changes:
Remaining true user decisions:
Next: active workflow owner

Return to the calling owner. For a Requirements call, that owner alone assembles the Candidate Requirement Set. For a Plan or lane call, return formulation or boundary deltas there; flag product-scope changes for Requirements routing.

If no material gap survives, name the dimensions and evidence inspected and say that no delta was found; do not invent issues to prove value. Omit empty headings. Read references/formulation-grilling.md for large, multi-surface, or multi-round uncertainty.

Red Flags

• Using grilling only to attack a proposal or produce a pass/fail verdict.
• Restating the input without identifying what changed.
• Listing generic risks without tying them to this formulation and its effects.
• Avoiding hard follow-up questions because the first answer sounds plausible.
• Stopping at a first-order consequence when it exposes another important gap.
• Returning questions without a resolver, disposition, and requirement impact.
• Asking the user before source- and agent-answerable gaps are exhausted.
• Inventing speculative gaps without a causal link or evidence need.
• Pruning a plausible branch before it is formulated enough for convergence.
• Publishing an authoritative Candidate Requirement Set instead of returning deltas to the calling owner.
• Calling discovered candidates the final Requirements Baseline.