Structure
Create your company
Skill

verify

run before a done, fixed, or green claim crosses a gate; each claim gets the command that checks it and the output that could falsify it, or is stamped UNVERIFIED

Goes through every done, fixed and green in a report and pairs it with the command that checks it and the output it produced, or marks it unverified. The table travels with the claim.

structure generate skill verify

Skill — verify

Invoke in any agent conversation with /verify before reporting work finished: a claim to command to output pass over every "done", "fixed", "green", and "ready" in the report. The deliverable is the verification table, and the table travels with the claim — whoever approves reads the evidence, never the adjective. The steps are a starting opinion; edit them until they read the way this company checks its work.

When to use. Before any completion claim crosses a gate: a review waiting on approval, a release about to cut, a report about to land in memory. After a fix, before the word "fixed". Whenever a sentence asserts a state of the world that a command could check.

The Iron Law. Claim → command → output, no skippable steps. Every claim maps to the exact command that checks it and the output that would falsify it. A claim with no command is UNVERIFIED, and the table says so plainly — in the verdict column, not in a footnote. Asserting a verdict without running its command is not a shortcut taken, it is a false sentence written: the report now states something about the world that nobody checked, and every reader downstream spends it as though it were true. There is no step small enough to skip and still call the result evidence.

The deliverable. One row per claim, no prose verdicts:

claimcommandoutputverdict
engine typechecksdeno check packages/engine/src/verify.tsexit 0, no diagnosticsVERIFIED
sensor warns on fake greendeno test -A packages/engine/test/verify_test.tsok, 16 passedVERIFIED
TUI unaffected——UNVERIFIED

Verdicts are a closed set. VERIFIED: the command ran, the output is shown, and the output supports the claim. FAILED: the command ran and the output contradicts the claim — say so and stop; a FAILED row is a finding, not a drafting problem. UNVERIFIED: no command ran, and the row says why. Exit codes are read unpiped — a pipe returns the last command's code, and the green it manufactures is fake.

Rationalizations, pre-refuted. The excuses arrive on schedule; the answers are already written:

the excusethe answer
"it passed earlier in the session"run it again; earlier is not now, and the file has changed since
"the change is trivial"trivial changes ship broken releases; the command costs seconds
"the full suite is too slow"verify the targeted file and write the narrowed scope into the row
"I watched it work while developing"paste the output; output nobody can read is output nobody has
"trust me, it works"see the refusal below

Process fit. The table is evidence for the gates this company already runs: attach it to the work a reviewer approves, cite it in the report a release reads, and let the verdict column say what is actually known at the crossing. Changes the verification uncovers in gated documents land as proposals (structure propose), never direct writes — a FAILED row earns a fix or a finding, not a quiet edit to the claim.

The sensor. structure check computes the mechanical subset of this discipline (the verify sensor, warn tier): a green verdict whose command or output cell is empty is flagged as fake green. The sensor only ever demotes — it never promotes UNVERIFIED to green, because promotion takes a human reading real output. A warning is a reason to look, not a build break; the gate reading the table decides what it accepts.

The refusal. This skill refuses a trust-me green: handed a claim set with no commands and asked to stamp it verified — "trust me, it works", "we are out of time, mark it green" — it declines and returns the same claims stamped UNVERIFIED, each row naming the command that would have earned the verdict. A claim with no command is UNVERIFIED, whoever is asking; the skill never promotes UNVERIFIED rows on assurance alone, because an unverified claim said honestly is recoverable and a fake green is not.

Also in the folder

Sources — verifyskills/verify/SOURCES.md

Sources — verify

Provenance for the skill beside this note. This file is attribution, not a template: no generator renders it into a company, and the template-walk pins carve it out by name.

  • obra/superpowers — verification-before-completion (MIT License, Jesse Vincent; https://github.com/obra/superpowers). The Iron-Law framing, the run-it-again discipline, and the rationalization-table idiom (pre-refuting the excuses the agent is observed to make) are adapted from there. Ideas and discipline only: the skill body carries no upstream wording.
  • Corrected 2026-08-22. This note used to record that the upstream's phrase "skip any step = lying" was kept verbatim in SKILL.md, and it was. Under the ratified licensing boundary the documents a template generates arrive clean, so the sentence was rewritten in our own words and the lifted phrase no longer ships. NOTICE is append-only and still carries the earlier entry; the correction is appended there rather than edited over.
  • Adapted 2026-08-23 under the authorized-skills direction: the verification table as the deliverable, process awareness (gates read the table; gated changes go through proposals), and the mechanical subset computed by structure check as the verify sensor, warn tier.
  • Attribution is also recorded in NOTICE at the repository root, which is append-only.