Kim OlsenAI consultant in Copenhagen. I set up Claude for sales and bid teams.
The cite check that looks up every promise in your tender before it goes out.
A figure rounded up. A requirement answered with the wrong number. A certification that is still in progress. These are the promises that become contract terms. The skill makes Claude check each one against the annex it cites, and the test pack lets you see what your own setup catches.
What you get.
- The cite-check skill in English and Danish, ready for Claude Code or the Claude app
- The invented tender with 26 promises and its five annexes, with ten planted errors
- The answer key for every promise
- The script that runs the check with your own API key, and the script that scores the reply
- A README with the commands and what the check will not do
What the test showed.
58 of 60
planted errors found by Sonnet 5.5 over six runs, with the annexes attached.
0 of 60
found by Sonnet 5.5 with the tender alone. It wrote that it could not check without the sources.
4.17
the one error that slipped through: a true promise citing the wrong section. The README says how to cover it.
A test, with a person signing off.
The check compares the tender with the documents you give it. If an annex is itself wrong, the error slips through. It does not replace the person who signs the tender: it turns their read-through into minutes instead of an evening.
Read how the test was built and what it found: A tender with 10 planted errors: what Claude found with and without the annexes →