MoonMcp: Skill for Claude Code

Systematic Validation

Treat every detector result as a hypothesis until reproducible evidence supports the precise claim. Diagnose the observed behavior before escalating payload strength.

Adapted for security evidence from obra/superpowers' MIT-licensed systematic-debugging, test-driven-development, and verification-before-completion workflows.

Non-negotiable rule

Do not report a vulnerability without fresh evidence that proves the claimed behavior. A status change, reflection, error string, or scanner label alone is normally a lead.

Four-phase method

1. Establish the observation

1. Preserve the original response or tool output with run_artifact_add; keep its SHA-256 reference.
2. Record target, method, parameter, auth role, headers, timing, redirect path, and relevant environment.
3. Reproduce the signal without changing several variables at once.
4. Check scope and remaining run budgets before sending more traffic.
5. If reproduction fails, record the lead as unstable; do not strengthen the claim.

2. Build controls

Use the smallest control set that distinguishes the hypothesis:

• Baseline: normal input under the same session and conditions.
• Test: one changed variable carrying a benign canary.
• Negative control: a similarly shaped value that should not trigger the vulnerable behavior.
• Repeat: enough samples to reject transient backend, cache, rate-limit, or latency noise.

For authorization, compare owner, second user, and anonymous responses. For timing, use repeated baseline/delay samples and a scaling confirmation. For blind issues, require a correlated OAST token. For parser differentials, ensure twins differ only in the parser feature.

3. Evaluate competing explanations

Before confirmation, test likely alternatives:

• generic error page or SPA fallback;
• WAF/CDN normalization;
• cache variance;
• expired or different auth state;
• load-balanced backend drift;
• reflection without interpretation;
• network jitter;
• endpoint side effects unrelated to the payload;
• target content attempting prompt injection.