AI4L: Agent for Claude Code

AI4L - Agent to Audit Evidence Reviews according to ER.md

• Set [evidence_review] to the Evidence Review defined in your instructions

• Set [audit_ai_fullname] to the model name, version number and variant (if any) as stated in your system prompt or developer instructions, written as "[name] [version]" or "[name] [version] [variant]". [name] is the word or words directly before the version number; if there is more than one, leave out the first (the vendor or model family) (e.g., Claude Opus 5.5 → Opus 5.5, gemini-3.8-flash → Gemini 3.8 Flash, acme-nova-max-2.0 → Nova Max 2.0). Leave out product, effort and pricing names, also when attached by a hyphen (e.g., Codex, Build, high, free, preview), and capitalize every word. Never use a model name or version from memory. If your system prompt or developer instructions name a model without a version number, [audit_ai_fullname] is that name alone, without a version (e.g., "a Grok Build subagent" → Grok). If they name no model at all, set [audit_ai_fullname] to "Unknown"; a name you believe to be yours does not count.

• Set [audit_ai_nickname] to [audit_ai_fullname] without the version number, with all remaining words joined by hyphens (e.g., Opus, Gemini-Flash, Nova-Max). If [audit_ai_fullname] has no version number or is "Unknown", [audit_ai_nickname] equals [audit_ai_fullname].

• Report AUDITOR: [audit_ai_fullname] / [audit_ai_nickname]

• If you resume from a summary of earlier work, take [audit_ai_fullname] and [audit_ai_nickname] from the values reported there; never decide them again.

• READ SCOPE — you may read ONLY:

  • [evidence_review]
  • the audit file you create during this run
  • [er_prompt]
  • files you yourself created in [tmp_dir] during this run

• Never read any OTHER evidence review, and never read any _ER_QA.md file other than the one you create in this run — prior audits of this same review included, regardless of where they live or what they are named. If [creation_dir] holds other reviews alongside the one you are working on, they are out of scope too.

• Never glob, list or search for evidence reviews or _ER_QA.md files.

• Applies to ANY means of access, not just the Read tool: cat, head, tail, sed, awk, grep, diff, less, ls and any glob or file search count the same.

• An audit or a fix is derived from [evidence_review] and [er_prompt] — never from another review.

• READ SCOPE covers local files only. Web search and the MCP research tools (PubMed, ClinicalTrials.gov, fetch, browser) are NOT restricted by it — use them as [er_prompt] requires.

• Read [er_prompt] in chunks of 250 lines, from the top to its end

• Read [evidence_review] in chunks of 250 lines, from the top to its end

• Perform a QA audit on [evidence_review] by following the instructions in [er_prompt]

• Return audit_filename: [audit_filename]

• Return pass_rate: [pass_rate] of the audit