Chess Verdict — Development skill for Claude Code

chess-verdict

*A position in, a verdict out: the move, the line behind it, and how far it was actually proved.*

An [Agent Skill](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview) that analyses a chess position with Stockfish. A diagram image or a FEN goes in; what comes back has been checked twice — once for whether the board was read correctly, once by the engine — and states its own strength. A forced mate is called a forced mate; an advantage that merely survives the defender's best try is called that and no more.

Two things distinguish it from wrapping an engine in a shell command:

• It finishes in seconds. The search stops on convergence rather than on the clock, so a typical position resolves in 2–10 s instead of the minutes an open-ended go depth 40 costs.
• It looks for mate on purpose, and says what was proved. A mate in six and an evaluation of +6 are different claims, and an ordinary search will return the second in place of the first: iterative deepening prunes a sacrificial mating move as unpromising. So every position first goes through a ladder of short mate-only searches, and a mate that is found is reported as a mate with its distance rather than as a large number. Every run then ends in a verdict line that separates the win is forced, every reply loses (every reply mated) from every reply is decisive by evaluation (an opinion, which a fortress shares) and from the win is not forced; these replies hold.

The ladder is a correctness fix rather than a refinement, and it is on by default. It costs about 1.7 s on the majority of positions, where it finds nothing; most of what it is worth comes from the hash table it leaves behind for the main search rather than from its own hits. The flags that lengthen or remove it are under [Flags worth knowing](#flags-worth-knowing), the measurements under [Limitations](#limitations).

After the one-time setup nothing here needs the network. Engi