Kissinger

The story (Isaacson's Kissinger; the aide was Winston Lord): Lord drafted a memo, handed it up. Kissinger returned it without reading - "Is this the best you can do?" Lord redrafted. Same question. Again. After several rounds Lord finally said, "Yes - this is the best I can do." Only then did Kissinger read it. The question wasn't cruelty; it was a forcing function that surfaced the gap between "done" and "best" that Lord himself couldn't see until asked.

Agents have that gap too. Asked "is this the best you can do?", an honest agent usually finds it isn't - and improves. This skill codifies the question, with a bound so it exhausts quality without iterating forever.

When to run

At the end of every phase (in a conductor: after the phase's work is gated, before it advances to the audit). Pairs with rumsfeld: run Rumsfeld first (surface the latent / unknown-knowns), then Kissinger - asking whether the whole, including anything Rumsfeld surfaced, is truly the best you can do.

The loop

1. Ask, honestly: "Is this the best you can do?" Answer against the actual work, not pride in it.
2. Hunt for a MATERIAL improvement - one that changes correctness, coverage, clarity, robustness, risk, or user value. Cosmetic/subjective tweaks do NOT count (see the bound).
3. If a material improvement exists → revise, VALIDATE it (re-run the gate/tests), resubmit, and ask again.
4. Stop when you can truthfully say: "Yes - this is the best I can produce under the current requirements and constraints," backed by what actually improved across the passes.

The bound - root it in reality (NOT infinite polishing)

The goal is rigorously exhausted, not forever iterated. Stop - and give the honest declaration - when ANY of these holds:

• No material improvement remains - the only changes left are cosmetic, subjective, or below the bar that would change a decision. That IS "best under constraints."
• Diminishing returns - this pass's improvement was materially smaller than the last and no longer worth a cycle.
• Pass cap reached - hard cap of 3 passes by default. A phase rarely needs more; if it seems to, that's a signal the plan/scope is wrong, not the draft. At the cap without a truthful yes, STOP and surface: what's still improvable, why, and whether it's a scope problem.

Guard BOTH failure modes: a premature "yes" on pass 1 (laziness - the thing Kissinger caught) AND an endless "no" (perfectionism). Neither is honest, and the human reviewing the run afterward will catch either.

The honest declaration (required output)

When you stop, state:

• Verdict - "best under current constraints," or "stopped at cap / diminishing returns."
• What improved across the passes - one line each (evidence the question did real work, not a rubber stamp).
• Residual limitations - what a better version would need (more time, a lifted constraint, data you don't have). "Best" means exhausted within constraints, with the constraints named - never merely asserted.

What this is NOT

• Not a rubber stamp - "yes" on the first ask is suspect; show that you looked.
• Not perfectionism - cosmetic churn past the material bar is a bound violation.
• Not scope creep - improvements are to the CURRENT requirements, not new features (that's the plan's job, resolved via the grill).