Verify

██╗   ██╗███████╗██████╗ ██╗███████╗██╗   ██╗
██║   ██║██╔════╝██╔══██╗██║██╔════╝╚██╗ ██╔╝
██║   ██║█████╗  ██████╔╝██║█████╗   ╚████╔╝
╚██╗ ██╔╝██╔══╝  ██╔══██╗██║██╔══╝    ╚██╔╝
 ╚████╔╝ ███████╗██║  ██║██║██║        ██║
  ╚═══╝  ╚══════╝╚═╝  ╚═╝╚═╝╚═╝        ╚═╝

Verify before you claim

A success claim without evidence is a false report, not efficiency. This skill is the gate between "I think it works" and "it works."

The gate

Before any completion claim, you must have run a check and read its output in this same response. No fresh evidence → no claim. Downgrade "it works" to "it should work — verifying now," then go get the evidence.

Claim → required proof

| Claim | What counts as proof (nothing less) | |---|---| | "Tests pass" | The runner's output, this turn — exit 0 and a non-zero test count with no unexpected skips (a run that collected 0 tests is a failure, not a pass) | | "The build works" | The build command's output, exit 0 | | "The bug is fixed" | The reproducing test, now green | | "Typecheck/lint clean" | The tool's output, not your reading of the code | | "The subagent completed X" | The git diff / files on disk — not the agent's success message | | "The feature works" | The feature exercised (test, curl, or a rendered view) producing the expected result | | "The migration / destructive op is safe" | It ran on a copy or dry-run and row counts + spot-checks match — not that the script reads correctly |

For UI specifically: render the real view and actually look at it — capture a screenshot with a headless browser, then read the image back in (and/or assert on the DOM / computed styles / a visual-diff baseline). Producing a screenshot you never inspect is the same blind-ship failure this rule exists to stop. Don't infer from CSS.

Goal-driven verification

The success criteria should already exist from forge:plan / forge:principles #4 — each task had a verify: check. Run those checks. If criteria were never defined, define them now and then verify against them.

Anti-evasion

Rephrasing doesn't exempt the rule. "Should be working," "looks correct," "I've implemented it" are all the same unverified claim wearing different words. Either you have this-turn evidence or you don't.

Deterministic vs not

This skill is for deterministic work — run it once, read the output. But if a supposedly-deterministic check has ever flaked (timing, ordering, shared state), it isn't deterministic — re-run it N times or fix the flake before trusting green. For genuinely non-deterministic behavior (LLM features, agents, prompts), a single green run is luck — use forge:eval (pass@k / pass^k) instead of a one-shot check.

If you can't verify

Say so plainly: name what you couldn't check and why, and present the work as unverified. An honest "I couldn't run the e2e suite — here's what I did confirm" is worth more than a confident false "done."

Exit

Green with evidence → forge:ship.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vasu-devs
• Source: vasu-devs/Forge
• License: MIT

Install and usage instructions live in the source repository linked above.