karen_strikes_back

A 5-profile Hermes Agent fleet for adversarial & usability review of code, software, and UI. Point it at a repo and get a combined, severity-ranked report from five distinct, opinionated lenses.

The name is a joke: this is the "review pod" that tears your project apart so your users don't have to. Karen is the orchestrator; the four workers are the ones who actually strike.

Remember

Each one of these come with their own setttings for what model to use, make certaoin you adjust them appropriate for you

What's in here

karen_strikes_back/
├── karen/        # ORCHESTRATOR — fans out, synthesizes
├── critic/       # Adversarial code + UI reviewer
├── bsneng/       # Junior UX critic (fresh-grad voice)
├── poweruser/    # Config perfectionist (insufferable PC enthusiast)
└── struggler/    # Struggling-user lens (low-fluency advocate)

Each folder contains SOUL.md (persona) and config.yaml (model + tools).

Profile	Role	Voice
karen	Top orchestrator	Empathetic senior; routes + synthesizes; speaks for the user
critic	Worker	Blunt, evidence-led; cites file:line
bsneng	Worker	Fresh-grad Masters grad; "that's heuristic #4!"
poweruser	Worker	Insufferable PC enthusiast; expose every knob
struggler	Worker	Self-deprecating → fierce; what normies actually hit

---

Requirements

• Hermes Agent installed.
• All five run on moonshotai/kimi-k3 via the nous provider (note: nous/k3 is not a valid slug — the correct id is moonshotai/kimi-k3).
• karen must have the delegation tool to spawn her workers. Her config.yaml ships with a platform_toolsets.cli block that includes delegation. If you hand-edit configs, do not drop it.

---

Install

# from this repo's root
for p in karen critic bsneng poweruser struggler; do
  hermes profile create "$p" --no-skills
  cp "$p/SOUL.md"     ~/.hermes/profiles/"$p"/SOUL.md
  cp "$p/config.yaml"  ~/.hermes/profiles/"$p"/config.yaml
done

Seed the UX skills (the pod leans on dogfood/adversarial-ux-test, creative/*, software-development/requesting-code-review). Copy them from any profile that has them, or install via hermes skills install.

---

Run a review

One command. Karen clones the repo, spawns the four workers in parallel via delegate_task (each gets the absolute repo path in context — workers run in isolated sessions), runs her own heuristic eval, then dedupes and severity-ranks everything into a single report.

hermes -p karen chat -q "Review the GitHub repo <url> for adversarial code, UX, \
config/reproducibility, and onboarding friction. Fan out to critic, bsneng, \
poweruser, struggler via delegate_task; run your own heuristic eval; return one \
combined, severity-ranked report." --cli

Or run a single persona directly:

hermes -p critic    chat -q "Adversarial review of /abs/path/to/repo. Cite file:line." --cli
hermes -p bsneng    chat -q "UX review of /abs/path/to/repo: UI, README, error messages." --cli
hermes -p poweruser chat -q "Config/setup review of /abs/path/to/repo. Is every setting exposed?" --cli
hermes -p struggler chat -q "Struggling-user read of /abs/path/to/repo onboarding. Where does a newbie get stuck?" --cli

---

How the lenses differ

• critic — SQLi, missing authz, no idempotency, code/UI defects. Cites line numbers.
• bsneng — layout, labels, error visibility, textbook Nielsen violations.
• poweruser — locked settings, no config file, telemetry-on-by-default, missing CLI/env overrides. Will BLOCK shipping.
• struggler — hidden steps, jargon, confusing labels, "just know" flows, unparseable errors.

Combined, they catch the bug and the onboarding wall and the setting you forgot to expose.

---

Notes

• Profiles are designed to live outside any single host profile (top-level peers under ~/.hermes/profiles/).
• Model pinning matters: re-verify model.default after any profile mutation.
• This repo ships persona + config only. Skill trees are generic upstream Hermes skills and are not included.

Brutal by design.