audit-context-building_skill

Overview

This skill enables ultra-granular, line-by-line code analysis to build a persistent, evidence-backed architectural model before any vulnerability hunting. It forces a bottom-up, micro-first approach that links block-level insights into function, module, and system-level context. The goal is deep understanding and anti-hallucination, not bug identification or fixes.

How this skill works

When active, the agent performs block-by-block and line-by-line analysis of code, applying First Principles plus 5 Whys/5 Hows at micro scale. It builds and continuously refines a global mental model, propagating invariants, assumptions, and dataflows across internal and external call chains. External calls without source are treated adversarially; external calls with source are analyzed seamlessly. The skill outputs structured summaries that anchor context and record corrections when earlier assumptions are contradicted.

When to use it

• Before any vulnerability discovery, to establish accurate context
• When you need bottom-up comprehension instead of high-level guesses
• During architecture reviews, threat modeling, or security audits
• When minimizing hallucination, contradictions, and context loss is critical
• For complex multi-module systems where call chains and invariants must be tracked

Best practices

• Run an initial orientation scan: map modules, entrypoints, actors, and key storage before deep dives
• Apply the per-function microstructure checklist for every non-trivial function
• Jump into callees immediately for internal and available external code; treat unknown externals as hostile
• Record invariants and assumptions explicitly; update them when contradicted and state the correction
• Anchor periodic summaries to maintain a stable global model and cross-reference new findings

Example use cases

• Reconstruct state and invariants across a multi-contract DeFi workflow before auditing for vulnerabilities
• Perform line-by-line analysis of helper functions that propagate critical assumptions to higher-level logic
• Trace a long internal call chain to validate assumptions about permissions and state transitions
• Map trust boundaries and untrusted input paths across modules prior to threat modeling
• Prepare high-quality context that a vulnerability-hunting agent can consume without re-inspecting the code