The 22 skills · Craft skill · Core Pack and Everything
AGENT SKILL
adversarial-review: what it makes your agent do, and what it refuses.
adversarial-review is the audit loop, written as a practice you can run. A maker cannot grade its own work, because it reads the intent rather than the bytes. Tuning a separate critic to be harsh is tractable, tuning a maker to be self critical is not, so this skill splits the roles and gives the critic its own eyes.
The job adversarial-review does
The loop has three roles that stay apart. Finders each attack from one angle and produce a reproduction rather than an opinion. Skeptics reproduce a finding from scratch before it is confirmed, because an unreproduced finding is a rumour. Sabotage breaks every fix and watches the guard fail, then restores it and watches it pass. Around those, four moves keep the rounds productive: refute by default so every check has a case that fails and a case that passes, weight each round toward the previous round's diff, recompute from the raw bytes rather than trusting the path under audit, and pivot to a restart when patching a broken approach costs more than replacing it. The exit condition is written down before the loop starts.
What adversarial-review refuses
A skill is worth what it stops as much as what it produces. These are the refusals written into adversarial-review, taken from the rules and limits in its own file.
- ✓A gate that cannot pass. A check with no passing case fails safe and protects nothing, which nobody notices until it matters.
- ✓A fix nobody broke. Adding a failing case is not proof. Revert the fix, watch the case go red, restore it, watch it go green.
- ✓A critic reading the maker's reasoning. Point it at the artifact, because a critic inside the maker's context starts agreeing with the maker's story.
- ✓A finding taken on trust. Reproduce it from scratch or it stays a rumour.
10 lines from the adversarial-review skill file, verbatim
From skills/adversarial-review/SKILL.md, unedited. Kymira skills are plain text you can
read in full before you install them, which is the point of shipping a method rather than a binary.
Every check needs two cases: the one that FAILS on bad input and the one that PASSES on good input. A gate that cannot pass is as broken as one that cannot fail — it just fails safe, so nobody notices it protects nothing. Write both, or you have a green light wired to no sensor. ### 2. Sabotage, or it isn't a test Adding the failing case is not enough — prove it fails for the reason you think. Revert the fix, run the case, watch it go red; restore, watch it go green. Two things this catches that nothing else does:
The failure modes adversarial-review addresses
These entries in the catalogue of 51 observed failure modes name this skill as what addresses them. Every one was seen in a real build or in our own audits, and none was invented.
The tier adversarial-review ships in
| Tier | Price | What the tier holds | adversarial-review |
|---|---|---|---|
| Free | $0 | honest-dataviz and the eleven doctrine principles, published in full | Not in this tier |
| Core Pack | $199 | the thirteen craft skills, one-time purchase, lifetime updates | Included |
| Everything | $349 | all 22 skills, both libraries, the template bank, one-time purchase, lifetime updates | Included |
Both paid tiers are a one-time purchase with lifetime updates, and all sales are final: the files are plain text you keep the moment they arrive, so judge the quality on the free skill and the published doctrine first. The pricing section states exactly what each tier contains.
Installing adversarial-review
A skill is a folder with a SKILL.md file in it. Installing one is copying that folder where
your agent looks for skills, with no build step and no dependencies.
mkdir -p ~/.claude/skills cp -R Kymira/skills/adversarial-review ~/.claude/skills/
Or open the unzipped pack in Claude Code and say install Kymira, which installs every skill in your tier, or double-click the installer for your system. Cursor reads the same skill as a rule file: copy adversarial-review.mdc into .cursor/rules/. Skills are read at startup, so start a new session and ask your agent which skills it has available.
Not for
Rounds are not correctness. The defect rate falls and does not reach zero, and after ten rounds an eleventh still confirmed fresh findings. Coverage is bounded by the lenses someone thought to write, a reproduction confirms one defect rather than its whole class, and a passing sabotage proves one guard rather than the system.
Last updated 2026-08-09 · Dylan, founder · one of 22 packaged skills, each one a method written down, not a product tour.