Skill Kit
Skill Kit wires the engineering literature into every stage of a Claude Code pipeline.
Most LLM-assisted workflows treat each feature as an isolated prompt: shape it, ship it, forget it. Two opportunities get left on the table: each feature could leave the system smarter about the codebase, and the pipeline's decisions could come from engineering literature, where today they come from memory and training data. Skill Kit wires both in.
Plumbing shippable skills is a well-trodden pattern. The interesting choice was wiring each stage to reach into distilled book knowledge instead of its own invented conventions, and then making the citations checkable. Every SKILL.md declares its sources in frontmatter, split into primary and secondary. /triage-issue reaches for Meadows' systems archetypes and Zeller's systematic debugging when a bug hits. /shape works off Gause and Weinberg's requirements elicitation and Duke's decision-quality-vs-outcome-quality split. /pre-merge reviews module depth against Ousterhout and reads its own findings through Kahneman's work on noise in judgment. The landing page's canon is generated from those frontmatter blocks by a script, so a book the pipeline stops citing falls off the page. The alternative was a hand-kept reading list, which is what most prompt libraries have: their "best practice" is whatever the author could remember writing down.
Skill Kit ships 33 skills: nine primary pipeline stages from /shape to /closeout, five invoked helpers like /tdd and /triage-issue, fifteen side routes with /qa and /prototype among them, and four infrastructure skills for hooks and guardrails. Each is a standalone SKILL.md with explicit preconditions, installable as a pack. State lives in GitHub: PRDs, slices, QA bugs, and lineage are issues and PRs. A review stamp on the PR is re-read at merge, so an approval that predates the latest diff cannot land. /pre-merge runs an eleven-dimension architecture review with an author mode, a reviewer mode for PRs by other authors, and a loop mode that records every finding in a durable ledger on the PR. /lfg runs the whole path unattended, from a short shaping conversation to a stamped PR, and never merges. Alongside the main pack, content-to-skill converts PDFs and EPUBs into progressive-disclosure reference skills that populate the personal library the pipeline cites.
Without it, you're YOLO-ing: shipping whatever best practice you happened to remember and hoping it's the one that mattered. With it, compounding knowledge: the same mistake never gets made twice, and a rule the pipeline enforces can be traced to the book it came from.
The pipeline treats the books as dependencies.