Fable 5 Code Review

Fable 5 Code Review brings whole-PR understanding to your review workflow by loading the entire changeset — plus the surrounding files it touches — into Fable 5's 1M-token context window. Instead of flagging line-level nits in isolation, it reasons across modules: tracing how a renamed function ripples through callers, whether a new code path violates an invariant established three files away, and where a refactor has quietly drifted from the codebase's existing architecture. It's built for engineers reviewing large or cross-cutting PRs where the hard bugs live in the seams between files.

When to use it

• Reviewing a PR that spans many files or touches shared interfaces, schemas, or state machines.
• Auditing changes to authentication, authorization, input handling, or anything where a vulnerability follows data across function and module boundaries.
• Catching architectural drift — a feature that bypasses the repository layer, a direct DB call in a controller, an abstraction leak.
• Onboarding to an unfamiliar codebase where you need the why behind a change, not just a list of diffs.

How it works

The skill collects the diff (via git diff against your base branch) and resolves the dependency neighborhood of every changed symbol, then submits the assembled context to Fable 5 for a single coherent reasoning pass. Output is grouped by concern — correctness, security, architecture — and each finding cites the specific files and lines involved along with an explanation of the failure mode and blast radius.

curl -fsSL --create-dirs -o ~/.claude/skills/fable-5-code-review/SKILL.md https://claudeskil.com/skills/fable-5-code-review/SKILL.md

# Review the current branch against main
claude review --skill fable-5-code-review --base main

# Focus on security data-flow analysis, post inline on the PR
claude review --skill fable-5-code-review --focus security --comment

You can scope a review with --base <ref> and narrow attention with --focus correctness|security|architecture. Add --comment to publish findings as inline PR comments.

Tips & gotchas

• Large context is a budget, not free lunch. If a diff plus its neighborhood exceeds ~1M tokens, the skill prioritizes changed files and their direct dependents — pull vendored or generated directories out of scope with a .reviewignore to keep signal high.
• Run it against a meaningful --base. Reviewing against HEAD~1 only sees the last commit; use the actual branch point for whole-PR reasoning.
• Architecture findings reflect patterns it infers from your repo. If you have explicit conventions, reference a CONTRIBUTING or ADR file so drift detection anchors to your intent rather than the statistical majority.

FAQ

Does it modify my code? No. Review is read-only and emits findings. Use a separate fix workflow if you want patches applied.

Will it replace my linter or test suite? No — it complements them. It targets semantic and cross-file issues that static linters and unit tests structurally cannot see, like invariant violations spanning modules.

Can it review private repos without sending the whole repo? It sends only the diff and the dependency neighborhood of changed code, not the entire repository, and respects .reviewignore exclusions.