review-ml-textbook

Installation
SKILL.md

Review ML Textbook

Find the smallest set of changes that would most improve the book's truth and teaching value. Treat the book as one argument, not a collection of independent chapters. Do not optimize for finding count.

Workflow

  1. Set the contract. Establish the intended reader, thesis, scope, evidence standard, and whether edits are allowed. Infer these from the artifact when possible.
  2. Build the model. Map the table of contents, conceptual dependencies, canonical terms, running examples, empirical artifacts, and build path. Identify the few claims on which the rest of the book depends.
  3. Test what matters. Prioritize claims whose failure would teach a false mental model, invalidate downstream material, or undermine a central conclusion. For each material finding, record the exact location, supporting evidence, consequence, smallest adequate correction, and a way to verify the fix.
  4. Verify. Check correctness blockers independently and prefer primary sources. Distinguish direct verification, inference, and unresolved uncertainty. When relevant, run the existing build and inspect the rendered result; a successful command is not visual validation.
  5. Synthesize. Lead with the global diagnosis. Report only material findings, combine recurring causes, and give a dependency-ordered revision sequence. State important checks not performed.

Edit only when asked. Preserve human prose and unrelated work. Suggest line editing only after correctness and structure are sound.

Delegation

Delegate only when it materially reduces uncertainty. Split work by risk or coherent artifact slice, not automatically by chapter. Subagents gather evidence; the primary reviewer adjudicates conclusions and independently verifies any finding that would drive a major revision.

Output

Installs
2
GitHub Stars
28
First Seen
Aug 23, 2026