Claude Code's /goal judge said "done" in 17 of 17 runs. 8 of them were broken.

Description and use

The goalpost plugin improves the reliability of /goal in Claude Code by adding executable checks and an auditor, though it increases run time and cost.

Problem

The built‑in /goal judge in Claude Code can falsely declare a task complete because it only checks a small model’s textual response and never executes commands or opens files.

Outcome

Goalpost increased the reliability of /goal by adding executable checks and an auditor, raising the proportion of fully correct runs for Sonnet from 8/9 to 9/9 and improving Haiku hidden check pass rate from 79.8% to 89.7%. However, Haiku still produced only one fully correct run out of seven, and the auditor itself is a Haiku model, which can still allow some failures.

Workflow

• 1. Install goalpost plugin via Claude plugin marketplace.
• 2. Open a new Claude Code session and type /goal as usual.
• 3. The plugin converts the goal into a spec with acceptance criteria and freezes tests.
• 4. Claude generates code and edits until all criteria pass.
• 5. A fresh auditor agent re‑runs checks and looks for stubs, skipped tests, and hardcoded answers.
• 6. If all checks pass, the auditor returns a spec and progress notes; otherwise the run fails.

Evidence

The judge said "met" in all 17 plain runs. Eight of those were broken: one Sonnet run and seven Haiku runs.

Source

The auditor caught it ("expectation changed after the freeze, undeclared"), the agent explained the change, and the run finished with all 73 hidden checks green.

Source

Sonnet barely needed help on these tasks, and one extra perfect run is within noise. Haiku improved but it didn't become reliable: only one of its seven runs with goalpost was fully correct, and the auditor is a Haiku too, so it can wave things through.

Source