Adversarialize

Perform the given task by spawning a worker agent, then an adversary agent to critique the worker's output, then a reconciler agent to apply the critiques you accept. You orchestrate and judge.

Arguments: $ARGUMENTS

Workflow

1. Take the task from the arguments. With no arguments, take the most recent task under discussion in the conversation. If there is no such task, ask the user what to do and stop
2. Spawn a worker agent with the task verbatim plus any conversation context it needs. Instruct it to perform the task and return what it changed and the reasoning behind non-obvious decisions
   • If the task is a slash command, instruct the worker to read that skill's SKILL.md (under ~/.claude/skills/ or the project's .claude/skills/), run any !-fenced context blocks as commands, and follow the skill with the given arguments
3. Spawn an adversary agent with the original task and the worker's report. Instruct it to:
   • Inspect the worker's output (diff, files, artifacts), not just the report
   • Make no changes: read, run read-only checks, and report
   • Try to refute the work: bugs, gaps, unmet requirements, and unjustified decisions
   • Return a numbered list of findings, each with location, problem, and suggested fix. Return "no findings" if the work is sound
4. Evaluate each finding yourself. Accept a finding only if:
   • It's correct. Verify uncertain claims by reading the relevant code or output
   • Fixing it serves the original task rather than the adversary's taste
5. If you accept no findings, report the worker's result plus the rejected findings with reasons, and stop
6. Spawn a reconciler agent with the worker's report and only the accepted findings. Instruct it to apply exactly those fixes
7. Report the worker's result, each finding as accepted or rejected with the reason, and what the reconciler changed

Guidelines

• Run the agents sequentially
• Don't perform the task, the critique, or the fixes yourself. Read only to verify adversary claims in step 4
• Keep the adversary independent by passing it the task and the worker's report, not your own opinion of the work
• Findings about the task statement itself (ambiguous or contradictory requirements) go to the user, not the reconciler
• If the reconciler reports changes beyond the accepted findings, spawn it again to revert the extras before reporting