For any non-trivial work — bulk edits, a migration, an analysis, a check — build the tool that does it or proves it instead of doing it by hand. A codemod, a script, a generator, or a skill the subagents follow. The result is then a by-product of something re-runnable.
The second clause is the load-bearing one and is what separates this from ordinary automation advice: the tool is the artifact a reviewer can rerun. The point is not that the agent works faster. It is that a reviewer who does not trust the agent’s claim can re-derive the result themselves, which converts an assertion into evidence. That makes this the constructive complement of pattern-evidence-before-claims: one says a claim needs proof, this one says build the thing that produces the proof.
It also answers a failure mode specific to agents. An agent that edits 200 call sites by hand produces 200 opportunities to silently diverge and no way to check; an agent that writes the codemod produces one artifact whose correctness is inspectable and whose application is uniform. pstack states it as a principle explicitly scoped to any non-trivial work, not just bulk work, closing the loophole where “this is only a few files” defeats it.
Applied by (backlinks)
pstack:
- pstack-principle-build-the-lever — the principle statement, with the reviewer-rerun clause.
- pstack-create-verification-skill — generates a per-repository harness that drives the real app, so behavior can be proven rather than asserted.
- pstack-blast-radius — rung 4 of its evidence ladder is a script that calls the real code and fails loud, shipped in the writeup.
- pstack-show-me-your-work — prefers evidence produced by committed scripts over hand-made one-offs.
See Also
- pattern-evidence-before-claims — the claim-side complement; this is the means of producing the evidence.
- pattern-deterministic-gates — what the lever often becomes once the lesson is worth enforcing.
- pattern-knowledge-compounding — a lever kept and reused is how tooling compounds.