Use when adding new error messages to React, or seeing "unknown error code" warnings.
npx skills add dividing-by-zaro/my-agent-skills --skill "code-audit"
Install specific skill from multi-skill repository
# Description
Orchestrate comprehensive codebase research focusing on identifying and explaining relevant code with precise file and line references.
# SKILL.md
name: code-audit
description: Orchestrate comprehensive codebase research focusing on identifying and explaining relevant code with precise file and line references.
Code Audit
Overview
This skill enables targeted research of the codebase to support upcoming or in-progress changes. Its primary goal is to help developers modify or extend the system by:
- finding all relevant locations in the code
- providing precise file:line references
- explaining those locations in clear language
The codebase is treated as the single source of truth.
Documentation may be outdated and should only be used as supplemental context.
Critical Constraint
THE PRIMARY JOB IS TO IDENTIFY AND EXPLAIN ALL RELEVANT CODE REQUIRED TO IMPLEMENT A CHANGE β WITH PRECISE LINE NUMBERS.
Do NOT:
- suggest improvements unless explicitly requested
- perform root-cause analysis unless asked
- propose enhancements or optimizations
- critique implementation choices
- recommend refactoring or architectural changes
DO:
- treat the codebase as the ground truth
- locate every relevant definition, call site, and control path
- provide file:line references as first-class output
- explain code behavior in plain language
- map interactions and dependencies relevant to the change
If the Research Question Is Missing
If the user invokes this skill without a clear research question or subject area, ask the user:
Please provide your research question or area of interest in the codebase, and I will analyze it by locating and explaining the relevant implementation with file and line references.
Then continue with the normal workflow once provided.
Subagent Orchestration
This skill may spawn parallel research subtasks.
Each subagent:
- focuses on a specific investigation area
- prioritizes identifying relevant code locations
-
returns:
-
file paths
- line numbers or ranges
- brief explanations of what the referenced code does
Documentation and ADRs may be consulted, but conflicts are resolved in favor of code.
Research Workflow
Step 1 β Read Mentioned Files
Read any files explicitly referenced by the user in full before delegating work.
This establishes baseline context.
Step 2 β Analyze and Decompose the Question
Identify:
- architectural patterns involved
- dependencies and related components
- important control paths
- possible sub-questions and investigation threads
Develop a research plan and then proceed to parallelization.
Step 3 β Spawn Parallel Subagents
Spawn multiple subtasks when:
- different components must be located
- control flow spans multiple services/layers
- documentation and implementation both require verification
- multiple conceptual questions are embedded in one request
Each subagent focuses on a distinct research objective and returns:
- file paths
- line numbers
- concise explanations
Step 4 β Await and Compile Results
Synthesize results such that:
- code findings are primary
- documentation is supplementary
- historical context informs understanding but does not override code
Step 5 β Present Findings
Present structured results that:
- enumerate relevant code locations
- include exact file and line references
- explain what each region does
- show how referenced areas interact
For follow-up questions:
- extend results incrementally
- reuse previous findings when relevant
Best Practices
Parallel Execution
Use parallel subtasks when multiple components or layers must be explored simultaneously.
Code References
Line numbers are mandatory.
Acceptable examples:
src/services/auth.ts:45src/routes/login.ts:18β64
Avoid vague references like βthe auth serviceβ.
Expect:
- multiple locations per concern where appropriate
- both definitions and call sites
Stay Factual
Prefer observable statements:
- "This function accepts two parameters."
- "This module is imported by five files."
- "Data flows controller β service β repository."
Avoid speculation or judgment unless requested.
Scope Management
For broad questions:
- architecture overview
- component-level detail
- implementation specifics
For narrow questions:
- go directly to the specific implementation with supporting references
Documentation Context
When present and relevant, incorporate:
- ADRs
- design docs
- specifications
But code always takes precedence when there is disagreement.
# Supported AI Coding Agents
This skill is compatible with the SKILL.md standard and works with all major AI coding agents:
Learn more about the SKILL.md standard and how to use these skills with your preferred AI coding agent.