AI4L - Main Skill for Evidence Review Creation and Auditing using @AGENTS

General Rules

• Parse the user's input to determine which sub-command to execute

• Set [args] to $ARGUMENTS

• Note the [start_time] when beginning any command, and report the [time_taken] when done

• All generated results go in [creation_dir] as .md files

• Do not edit or modify any files outside [creation_dir]

• Full lines formatted as line comments must be ignored when processing commands.

Resolving the target ER

• If no [args] are given, set [target_er] to the newest _ER.md file in [creation_dir]
• Otherwise, if [args] names an existing _ER.md file in [creation_dir] (with or without the directory), set [target_er] to that file's bare filename, without any directory
• Otherwise treat [args] as an intervention: list the _ER.md files in [creation_dir] whose filename starts with [args] converted the way 6.25 of [er_prompt] converts short_topic (lowercase, spaces and dashes to underscores, other characters removed), and set [target_er] to the newest of them by the date in the filename
• If nothing matches, report No ER found for: [args] and exit

COMMAND: create {topic}

Create an evidence review (ER) using the @er-creator agent.

• If no [args] are given {

  • Report: usage: /er create {topic}
  • exit } otherwise { set [topic] to [args] }

• Report: create: [topic]

• @er-creator: [topic]

• Wait until the agent finishes

• Report: filename: [filename]

COMMAND: audit {er}

Audit an ER using the @er-auditor agent.

• Resolve [target_er] from [args] (see "Resolving the target ER")

• Report: audit: [target_er]

• @er-auditor: [target_er]

• Wait until the agent finishes and returns the result

• Report: target_er: [target_er]

• Report: pass_rate: [pass_rate]

• Report: audit_filename: [audit_filename]

COMMAND: fix {er}

Audit and fix an ER using the @er-fixer agent.

• Resolve [target_er] from [args] (see "Resolving the target ER")

• Report: fix: [target_er]

• @er-fixer: [target_er]

• Wait until the agent finishes and returns the result

• Report: target_er: [target_er]

• Report: pass_rate: [pass_rate]

• Report: audit_filename: [audit_filename]

COMMAND: combine {er}

Create a final _ER_QA.md file from all audits

• Resolve [target_er] from [args] (see "Resolving the target ER")

• Report: combine: [target_er]

• @er-combiner: [target_er]

• Wait until the agent finishes and returns the result

• Set [new_qa_filename] to the filename the agent returns in its QA file: line

• Report QA file: [new_qa_filename]

COMMAND: iterate {er}

Loops audit/fix cycles up to [max_audits] times until [needed_passes] show 100% pass rate.

• Resolve [target_er] from [args] (see "Resolving the target ER")

• Report: iterate: [target_er]

Initialize {

• Set [iteration] = 0, [consecutive_passes] = 0 }

Loop while [iteration] < [max_audits] and [consecutive_passes] < [needed_passes] {

• @er-fixer: [target_er]

• Wait until the agent finishes and returns the result

• If the [pass_rate] is 100%, increment [consecutive_passes]; otherwise reset to 0

• Increment [iteration]

• Report: "Iteration [iteration]: Pass rate = [pass_rate] ([consecutive_passes]/[needed_passes] consecutive passes needed)" }

• @er-combiner: [target_er]

• Wait until the agent finishes and returns the result

• Set [new_qa_filename] to the filename the agent returns in its QA file: line

• Report QA file: [new_qa_filename]

If [consecutive_passes] >= [needed_passes] {

• Return: status: success } else {

• Return: status: failed }

• Return: target_er: [target_er]

• Return: iterations: [iteration]

COMMAND: full {topic}

A create and multi-pass audit workflow.

• If no [args] are given {

  • Report: usage: /er full {topic}
  • exit }

• Execute the "create" command

• Execute the "iterate" command with [filename] returned by "create" as [args]

COMMAND: compare {intervention}

Compares all ERs for a given intervention, typically from different AI models or versions, to determine which is strongest based on content quality and the latest QA audit results.

• If no [args] are given {

  • Set [intervention] to intervention in the frontmatter of the newest ER in [creation_dir] } else {
  • Set [intervention] to [args] }

• Report: compare: [intervention]

• Compare all ERs in [creation_dir] with a similar [intervention] in their frontmatter

  • Compare the quality of the content
  • Be detailed
  • Take into account the latest "_ER_QA.md" for each.

• Present a clear recommendation of which ER is strongest and why.