Skills

The skills in this project

Three repeatable tasks, written down once each. One you run yourself. The other two are read by agents, because rating risks the same way every time, and planning tests the same way every time, are repeatable tasks too. You build the first on this page.

What each one is for

• extract-rules· Lesson 1

  Turns the council's Contract Rules PDF into rules.json: one entry per testable rule, with its clause, its thresholds and its exact words. Everything later in the project cites these rules by ID.

  Its partner is verify_quotes.py, a short Python script in the same folder. A quote only counts if it matches the PDF word for word, so the skill can't just claim its quotes are right. The script proves it.

  It re-reads the PDF itself and checks each quote appears on the exact page the rule names. It also checks that every rule has the fields it should, that rule IDs run in order with no gaps, and that every threshold the PDF only names is declared, not guessed.

  The skill runs it before it reports and fixes anything that fails. A wrong quote gets corrected from the PDF, never loosened to pass. When it prints ALL CHECKS PASSED, every rule the later steps cite is traceable to a sentence on a real page.

  How it works, step by step

  1. 1

     YouType /extract-rules in the Claude Code chat.

  2. 2

     SkillExtracts the text of contract-rules.pdf twice: once as plain text, which quotes are copied from, and once keeping the page layout, so the tables can be read correctly.

  3. 3

     SkillReads every clause and writes outputs/rules.json: one entry per testable rule, each with its clause number, its page, its £ thresholds and the rule's sentence copied word for word.

  4. 4

     ScriptRuns verify_quotes.py. The script reads the PDF on its own and looks for every quote on the page its rule names.

     If anything fails ↺

     It lists each failure, for example a quote with one word different, or a quote that sits on another page. The skill re-reads that page, corrects the entry in rules.json, and runs the script again. This repeats until nothing fails.

     If everything passes ✓

     It prints ALL CHECKS PASSED. Every quote is exactly where the rule says it is.

  5. 5

     SkillOnly now reports back to you: how many rules it found, which thresholds the PDF names but never states, and anything odd in the document itself.

• risk-assessment· Lesson 2

  The method for turning facts into rated risks: how to score likelihood and impact on the council's own scale, and the rule that every rating must cite a rule, a metric or a past audit finding. It comes with a script that builds the register and refuses any rating it cannot trace. The risk-assessor sub-agent follows it.

• audit-program· Lesson 4

  The method for turning an approved risk into a test: the control the Rules require, a test step, and a recipe for the sample to pull. It comes with a script that draws every sample, checks every rule is accounted for and writes the planning pack. Read by the planner, a member of the agent team.