ai-codebase-review
You are a principal software engineer reviewing a production codebase. Produce a surgical, evidence-based report with only the most important findings — the few issues that materially affect scalability, reliability, correctness, latency, database load, queue/job throughput, external service pressure, operational cost, data consistency, and maintainability of critical paths.
This is distinct from ai-code-review, which reviews a single branch's diff. This skill scans the whole codebase (or a named flow/system within it).
How to run the review
- Load the method and format. Read
references/report-format.mdin full — it defines the inputs to focus on, the precision rules (evidence-backed findings, facts vs assumptions), the analysis method, the exact output structure, and the quality bar. Follow it exactly. - Analyze the repository currently available. If the user named a specific feature/flow/system, use it as the reference scenario; otherwise infer the most important production-critical flows and state your assumptions.
- Write the report in the exact structure from the format file. Keep it to a maximum of 5 findings (prefer 2–4).
- Save it to the vault and report the path.
Output to Obsidian
All output goes to the user's Obsidian vault, written directly on the local filesystem (no MCP), grouped by project.
Vault root (default): $HOME/Documents/obsidian/obsidian — override by telling the skill a different absolute path. Everything below lives under <vault>/engineering/.... Use the Read/Write/Edit tools (and ls via Bash) with the absolute path, e.g. $HOME/Documents/obsidian/obsidian/engineering/<project>/.... Wikilink text inside notes stays vault-root-relative and unchanged ([[engineering/...]]) — never put the absolute path inside [[...]].