A fast agent can carry a wrong number straight into a decision. These skills put an argument in the way. The agent is told to push back on you, on the draft, and on its own answer before anyone acts.
Each refusal below quotes the skill that carries it. A test in this repository fails if the quote stops matching the skill.
No recommendation from a number nobody reproduced
A recommendation built on an unreproduced number inherits its error.
From make-the-call. The agent reproduces the result first, and diagnose-change tells it to reconcile the recalculation to the reported figure:
Recalculate the reported movement from the smallest sufficient source and reconcile it to the reported number.
No story from the aggregate alone
A rate can fall overall while every segment improves.
From the rigor rules. When a total moves, the agent separates change inside the segments from change in their weights. In the conversion case, a synthetic teaching case, that is the difference between "signup broke" and "traffic shifted."
No business fix from arithmetic
Recommend the smallest useful next observation or action; do not prescribe a business intervention solely from a descriptive decomposition.
From diagnose-change. A decomposition shows where the movement came from. It does not show why behavior changed.
No causal verb the design hasn't earned
Use causal language only when the design earns it.
From the rigor rules. "Caused," "drove," and "thanks to" need randomized assignment, or an identification strategy whose assumptions are named and testable.
No costly action on descriptive evidence
A costly or irreversible action needs evidence of cause.
From make-the-call. A descriptive finding can support a cheap, reversible step, such as a check or a small test. It can't carry a costly or irreversible one on its own.
No invented inputs
Never fill it with an estimate you made up.
From make-the-call. A cost, value, or threshold the inputs don't supply reads "unknown," with what the owner would need to provide.
No decision manufactured from a descriptive ask
This skill does not manufacture a decision.
From make-the-call. With no decision in view, it says so and stops.
No confident answer when the data can't give one
"Not enough data," "inconclusive," and "the question needs reframing" are complete findings.
From write-finding. The non-answer comes in the first clause, before the reason.
No result for a query that didn't run
Never fill a saved result with expected values.
From the rigor rules. A check that didn't run is labeled unexecuted.
No approval it can't give
The owner decides. A recommendation is a proposal, not an approval, and an agent never records one as accepted.
From make-the-call. Metric definitions work the same way: they stay proposed until the right person approves them.
What this doesn't promise
These are instructions to your agent, not enforcement. An agent can still break one, and a portable answer carries no guarantee because a skill produced it. The skills lab records one isolated agent on four synthetic tasks, which is a teaching record, not a benchmark.
Two things help catch a miss. analysis-review challenges the result as a separate step. The optional Engine adds runnable Checks, and even those establish only the properties they test.