aegis

aegis

!checklist init ${CLAUDE_SKILL_DIR} --force

Security is not the last check before launch; it is a line that runs through every part of the lifecycle. Treating it as something you add once the features work — bolt-on security — fails almost by construction, because the most damaging weaknesses live in architecture and design, where they cannot be patched in afterward, only rebuilt. The correct posture is the one this whole suite keeps returning to: shift security left (the DevSecOps idea), the same "catch it early, surface it cheap" instinct as quality-left and red-means-stop. This skill is the shield — the discipline of building a system secure-by-design — across six gated stages, and it will not advance past a GATE until the checklist tool clears it. Its adversarial counterpart, the spear that attacks what this builds to prove it holds, is the gungnir skill.

Security is where the agent era is most dangerous, because a vulnerability is the worst possible fit for how an agent works:

• The agent writes insecure code by default. Its training distribution is full of vulnerable patterns — string-concatenated SQL, eval, weak or homemade crypto, hard-coded keys — and it reproduces them, because insecure code works perfectly until someone attacks it.
• A vulnerability passes every green test. It is the absence of an attack so far, not a failure — so it sails through every functional test and every green dashboard, and a green-optimizing agent ships it believing the job is done. Security is the ultimate "green ≠ correct."
• The agent has no threat model and trusts its input. It does not spontaneously ask "who would attack this, and how"; it writes the happy path and treats input as friendly, which is injection and XSS.
• The agent invents crypto/auth and over-grants to make it work. It will happily generate a custom authentication scheme or widen a permission to *:* to clear an error, feeling none of the risk.

So the rule that governs this skill: security must be woven in and enforced at every stage, never trusted to instinct — because the most prolific contributor is an agent whose every default is the insecure one, and whose green light certifies nothing about whether the system can be attacked. The aim is not absolute security (which is impossible and infinitely costly) but risk management: identify what is most worth protecting and most likely to be attacked, and concentrate the defense there.

Discipline: finish every GATE before the next stage. GATEs are hard — never skip, batch past, or self-certify a stage you have not done. The checklist tool enforces the order; let it. Commands address stages by name.

Read references/agent-era-shifts.md first — it is the heart: what each security practice becomes once the code is written by something with no threat model, no felt risk, and a relentless pull toward whatever turns the light green. If $ARGUMENTS is a throwaway with no sensitive data and no exposure, this machinery is overkill — say so. The leanest sufficient defense for the system's real risk is the goal, not maximum ceremony.

Speak the user's language, or the risk gets accepted blind. This skill makes the user own real trade-offs only they can price: what data is sensitive, what an outage or breach costs, which compliance regime applies, how much defense a risk earns. Read their fluency and gloss a term on first use (threat model, STRIDE, least privilege, defense in depth, SAST/DAST/SCA, CIA, zero trust). A user who signs off on "this residual risk is acceptable" in words they never parsed has not accepted it — and the FRAME, MODEL, and OPERATE judgments this skill leans on them for are then hollow.