← ClaudeAtlas

adversarial-devlisted

Orchestrates a code task through two separate agents in an adversarial loop — an executor that implements and a strict evaluator (no write tools) that scores the result 0-10 and demands fixes. Use when the user wants a task "executed and evaluated by 2 agents", "adversarial dev", "generator/evaluator loop", higher-reliability implementation with independent review, or says "turn this into an adversarial workflow".
Ronnasayd/aipp · ★ 3 · AI & Automation · score 62
Install: claude install-skill Ronnasayd/aipp
# Adversarial Dev Runs one coding task through two independent agents — an **executor** that implements and an **evaluator** that only critiques — looping until the evaluator scores the result 8+/10, stagnation is detected, or 8 iterations are exhausted. The evaluator never writes code; it only reads, runs lint/tests on changed files, and scores. This mirrors GAN-style adversarial review: the same agent grading its own work is structurally biased, so grading must live in a separate, tool-restricted context. ## Instructions ### Step 0: Gather task input From the user's request extract: - **Objective**: what must be built/fixed. - **Acceptance criteria** (optional): if the user gave explicit criteria, use them verbatim. If absent, generate 3-6 concrete, checkable criteria yourself from the objective (functional behavior, edge cases, existing conventions in the repo). - **Difficulty**: classify once, now, 0-10 — easy (0-3), medium (4-7), hard (8-10) — based on scope (files touched, new logic vs. edit, ambiguity). This does not change the iteration cap; it's for your own calibration and for the log. Create a log file at `<scratchpad>/adversarial-dev/<task-slug>/log.md` (use the session scratchpad directory) with objective, criteria, and difficulty. Append to this file after every iteration — it is the durable record if the loop stops without success. If this run is one of several parallel invocations (see "Parallel usage" below), `<task-slug>` MUST include a unique id (e.g.