Skill: empirical-proof

Purpose

Eliminate hallucinated completion. Coding agents are pattern-completers; the pattern of "successful task" includes confident-sounding completion language ("✅ all tests pass") that the agent will produce regardless of whether the underlying claim is true. Empirical proof is the structural defence: the agent cannot complete the pattern without first pasting evidence the pattern is true.

Resolving the project's commands

Before pasting any "verification output", resolve the actual commands this project uses to validate and test — typically from the repo's contributor docs (e.g. an AGENTS.md, CONTRIBUTING.md, or the scripts block of a manifest). Some projects also run an extra check (dependency or architectural-rules validation) beyond build and test. If you can't find the command, or aren't sure which one a claim refers to, ask the user before running anything — guessing the command produces output that proves nothing.

Core rules

1. Don't assume success

Writing the code is 10% of the job; verifying it works in the current environment is the other 90%. An unrun command's output is a guess, so run it before claiming what it says.

2. Verbatim pasting

When recording your completion claims, paste the verbatim output. No paraphrasing, no summarising, no "✅ passing" — a summary is unfalsifiable, so it doesn't satisfy the gate. Use a fenced code block. Include the last two lines (or more if asked): the runner's summary plus its timing/exit conditions.