Your PR is a monster.
Five ways to understand it before you type “LGTM”.

Vue Vapor
Silk gzip decoder
Pinned source snapshots. Vue and React statistics exclude the two showcase artifact commits; their PRs live in our forks.
01Walkthrough
Start somewhere better than file one.
A diff has an order. It probably isn't the order you need. Follow a reading path that explains the change, then puts the relevant source beside it.
/walkthroughTry it: use Next to follow Silk's fast decoder from its guard rails into the code.
Published run, mobile layout fixed. Complete published walkthrough from the final merged source head.
02Reverse design
Read the design hiding in the diff.
Reconstruct the moving parts, the boundaries and the awkward cases. Architecture and sequence diagrams give you something to check against the source.
/reverse-designTry it: explore Vue's design documents and follow a diagram's source references.
Original published viewer. Complete published design viewer, with documents, diagrams and source references.
03Code map
Find the code behind the claim.
Thousands of files. A handful of useful starting points. Search the inventory, inspect a function, and follow the connections the analysis can support.
/code-mapTry it: inspect compileProgram, follow its connections, then use zoom and Fit files to explore this excerpt.
Real run, excerpt. Eight compiler files from a 3,669-file source change. The source and supported connections are real; the complete map is in the original run download.
04Review quiz
Find out what you missed.
Reading an explanation can feel suspiciously like understanding it. Answer questions about behavior and architecture, then check your reasoning against the evidence.
/review-quizTry it: choose an answer, check the feedback, and follow the source. Retakes are encouraged.
Original published viewer. Complete published quiz. Your score measures this exercise, not PR correctness.
05Comics
Remember the bit that bites.
Some ideas stick better as a small argument between functions. A source-backed comic makes one tricky constraint memorable enough to bring into the review.
/xkcdTry it: read the strip, then open its transcript to see where the metaphor stops.
Real run, excerpt. One of three original strips. Panel dialogue and artwork are preserved; the transcript links the metaphor to source.
We brought actual monsters.
A compiler import, a new renderer, and a decoder with an unsafe fast path. Here's what the skills helped make legible.
React Compiler
Where do you start with 3,669 files?
Use the map to find an entry point, then the walkthrough to connect it to the compiler pipeline.
Read the runVue Vapor
Follow a template all the way to the DOM.
Vue Vapor shows where design documents and a reading order make a cross-package change easier to understand.
Read the runSilk gzip decoder
Read the fast path without skipping the guard rails.
Only 18 changed files. Plenty to understand about buffer proofs, decoder handoffs and CPU-specific checksums.
Read the runBring your own monster.
Free, open-source skills. Your existing coding agent. Start with one format, or run /code-unpacked for all five.
npx skills add sandoche/agentic-code-unpackedThen bring a PR to your agent:
/walkthrough pr https://github.com/owner/repo/pull/123 --draftUsing Codex?
$walkthrough pr https://github.com/owner/repo/pull/123 --draftRequirements and supported agents
Node.js 22.23.4+ with npm. Start a new agent session after installation. First use downloads the rendering tools.
Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode.
Read the usage guideFair questions.
What am I installing?
Instructions and tools for the coding agent you already use. Install an individual skill, or code-unpacked for the complete five-format bundle. There's no separate Code Unpacked account.
Is this only for AI-written code?
No. Bring a PR, branch, repository, selected files or pasted code, whoever wrote it. AI can make the pile grow faster; understanding the pile is the useful part.
How is this different from a PR summary?
A summary tells you what changed. These formats give you a way through it: a reading order, reconstructed design, explorable connections, questions to answer, or a comic about one important constraint.
Can the explanations be wrong?
Yes. Your agent generates them. Use the source references to check important claims. Missing evidence, inferred intent and unsupported code should stay visible. The real runs show those boundaries too.
Does a good quiz score mean I should approve?
It means you understood the exercise. Correctness, tests, security and whether the change solves the right problem still need your judgment.
Where does my code go?
Your agent's configured providers and data-handling rules apply. Artifacts save under .code-unpacked/. Use --draft to keep delivery local. Posting to a PR requires an explicit request and authenticated access.
Can I open the results without the agent?
Yes. Finished HTML opens locally and works offline. Large outputs may arrive as a ZIP: extract it first. The website previews are selected from real runs; each links to its original result and complete bundle.
Is it free?
The software is free, open source and MIT licensed, including commercial use. Your agent, model provider and image-generation service may charge for usage.
What do comics and video need?
Comics need an available image-generation tool. Without one, the workflow can retain its storyboard and prompts. Optional video export needs FFmpeg; the interactive HTML does not.