# karen_strikes_back: a five-persona review pod that tears your repo apart

*2026-07-20*

> Five opinionated Hermes Agent profiles — an orchestrator plus four adversarial lenses — that review your code and UI in parallel and return one severity-ranked report.


Code review by one model is one opinion. Code review by five models with *different personalities, priorities, and axes to grind* finds things no single pass catches. That's the premise of `karen_strikes_back`: a fleet of five Hermes Agent profiles — the name is the joke, Karen is the orchestrator and her workers are the ones who actually strike.

Each worker is a distinct lens with its own `SOUL.md` persona and pinned model config. **critic** is blunt and evidence-led, citing `file:line` for every defect — SQLi, missing authz, broken idempotency. **bsneng** reviews like a fresh Masters grad who's read every Nielsen heuristic and will tell you which one you violated. **poweruser** is the insufferable PC enthusiast who checks whether every knob is exposed and will block shipping over telemetry-on-by-default. **struggler** is the self-deprecating advocate for users who bounce off jargon, hidden steps, and "just know" flows. Karen fans the repo out to all four in parallel via `delegate_task`, runs her own heuristic eval, then dedupes and severity-ranks everything into one combined report.

Features:

- One command reviews a whole GitHub repo across four adversarial dimensions at once
- Each persona runs in an isolated session with the absolute repo path in context
- Combined severity-ranked output — deduplicated across lenses, not four separate essays
- Personas are top-level peers under `~/.hermes/profiles/`, not tied to any single host profile
- Model pinning per profile (`moonshotai/kimi-k3` via the `nous` provider), adjustable to whatever you run
- Run any single persona directly when you only want one lens

*Brutal by design.* Source: [github.com/rahlquist/karen_strikes_back](https://github.com/rahlquist/karen_strikes_back)