Shipmates
All orders

Order

/harden

threat-model → remediate → re-review

Security-harden a surface — threat-model it, rank the findings, remediate the blockers, and re-review until every finding has a fix or an explicit accepted-risk note (and secrets/dependency scans are clean).

Take a surface and make it defensibly secure: a security-engineer threat-models it, findings are ranked and fixed, and the loop closes only when every finding is either remediated or carries an explicit, reasoned accepted-risk note — and the secrets/dependency scans are clean. The gate is "nothing Critical/High left unaddressed," not "looks fine."

Input ($ARGUMENTS): the surface to harden — a module, an endpoint/route, an auth or payment flow, a dependency set, or the whole app. If empty, ask which surface (don't silently pick the whole repo).

How to run it

Run it in Claude Code

/harden <what to harden — a module, an endpoint, an auth flow, or the whole app>

<angle brackets> = required · [square brackets] = optional

Step by step

The stages

Six stages, in order.

  1. Scope the attack surface

    Map what's in scope: the entry points (routes, CLI, message handlers), the trust boundaries, what data crosses them, and what an attacker controls. List the assets worth protecting. Keep the surface bounded to $ARGUMENTS — note explicitly what you're not covering so it isn't mistaken for cleared.

  2. Threat-model & find

    Crew: security-engineer (agent: security-engineer)

    Spawn the security-engineer to walk the surface with STRIDE across each trust boundary and review against OWASP fundamentals (broken access control / IDOR, injection, secrets & crypto, supply chain, secure defaults / defence in depth). Run the repo's existing SCA / secret scanners as corroboration — but reason about data flow, don't just collect lint. Return findings ranked Critical/High/Medium/Low, each with the exploit path (inputs → impact), exact location, and a specific fix.

  3. Triage

    Separate blockers (Critical/High, and any hardcoded secret / clear injection / broken authz) from nits (Low, defence-in-depth niceties). For each blocker: fix it, or — if it's a deliberate, justified risk — record an accepted-risk note (what, why it's acceptable, who'd own it). Nothing Critical/High may be silently dropped.

  4. Remediate

    Crew: senior-engineer (agent: senior-engineer)

    Spawn a senior-engineer with the exact blocker list; apply scoped fixes (parameterise the query, add the authz check, move the secret to config, pin/upgrade the dep) — no unrelated rewrites. Where a fix changes behaviour, add/adjust a test so it's covered. Rotate any exposed secret is a human action — flag it loudly; don't just delete it from the diff and consider it solved.

  5. Re-review

    Crew: security-engineer (agent: security-engineer, fresh pass)

    Re-run the security-engineer against the remediated surface: confirm each blocker is actually closed (not moved), no fix introduced a new hole, and the secret/dependency scans are clean. Loop Stages 3–4 up to MAX_ROUNDS; if blockers remain after that, STOP and escalate with the open findings — never declare a surface hardened while a Critical/High stands.

  6. Report (and PR, if MODE=fix-pr)

    Deliver: the threat model summary, the findings table (severity → status: fixed / accepted-risk / deferred), the remaining risk in plain words, and any human follow-ups (secret rotation, infra/config changes outside the repo). In fix-pr mode, open a CI-gated PR (reuse /ship-issue's worktree + CI gate + trailers) and file Medium/Low items as labelled follow-up issues.

Config

  • REVIEWER = security-engineer. BUILDER = senior-engineer. MAX_ROUNDS = 4.
  • MODE = report (default: produce a prioritised remediation report, fixing what's safe and in-scope) or fix-pr (apply remediations on a worktree branch and open a CI-gated PR, like /ship-issue). Infer from the request; when ambiguous, default to report.
  • Threat model / sensitivity = from the repo (what data it handles, what's exposed). Assume the repo is public and input is hostile.

Guardrails

  • Proportionate, not paranoid — harden to the real threat model; don't gold-plate a low-value surface.
  • Show the exploit path, not a scanner hit — a finding without a plausible attack isn't a blocker.
  • Never write a secret/token into a commit, PR, issue, or log; flag exposed secrets for rotation (deleting them from the diff does not un-leak them).
  • Every Critical/High ends as fixed or explicitly accepted — never silently dropped.
  • Bounded loop; escalate with open findings rather than rubber-stamping.
  • Drops into /ship-issue's acceptance board too: on a security-sensitive story the security-engineer is the gated reviewer. If the role doesn't resolve to a .claude/agents/*.md, fall back to general-purpose with the brief inlined and note it.

Where this lives

This page is generated from skills/harden/SKILL.md. The installer copies it to ~/.claude/skills/harden/SKILL.md for every project, or .claude/skills/harden/SKILL.md inside a single repo.

View skills/harden/SKILL.md on GitHub

Other orders

All orders