claude-howto: the folder numbers and the learning path disagree, and the tags stopped in June

The folder numbers and the recommended order are two different sequences

The ten modules are directories named 01-slash-commands through 10-cli, which invites the assumption that the numbering is the curriculum. It is not. The recommended learning path runs Slash Commands, Memory, Checkpoints, CLI Basics, Skills, Hooks, MCP, Subagents, Advanced Features, Plugins, which in directory terms is 01, 02, 08, 10, 03, 06, 05, 04, 09, 07. Skills is folder 03 but module 5. Checkpoints is folder 08 but module 3, placed third because it is treated as intermediate. MCP is folder 05 but module 7, after Hooks. The ordering is deliberate, since prerequisites come earlier, but the mismatch is real: a reader who types 03 into the sequence to follow module 3 gets Skills, not Checkpoints. Every module link in the tables carries its folder name, so the fix is to follow the table rather than the numbering, which is an extra step the layout does not make obvious.

The Beginner time estimate does not match the modules labelled Beginner

The level table promises a Beginner track of about 2.5 hours, Intermediate about 3.5 hours and Advanced about 5 hours. The module table breaks those into ten line items. Add up all ten and you get 690 minutes, or 11 hours 30 minutes, which sits sensibly inside the quoted 11 to 13 hours for the whole path. The per-level numbers are where it breaks down. Only two modules are labelled Beginner: Slash Commands at 30 minutes and Memory at 45 minutes, which is 1 hour 15 minutes against a quoted 2.5 hours. Adding the one module labelled Beginner+, CLI Basics at 30 minutes, still leaves you at 1 hour 45 minutes. The consequence for a reader is small but real: someone planning an afternoon against the 2.5 hour figure finishes early and concludes they skipped something, when in fact the level figure appears to bundle in material the module table assigns to a later level. Budget from the module table, not the level table.

The product is a set of files you copy, and every copy becomes yours to maintain

There is nothing to install from this repository. The quick path is a clone and five copy operations into your own project:

git clone https://github.com/luongnv89/claude-howto.git
cd claude-howto
mkdir -p /path/to/your-project/.claude/commands
cp 01-slash-commands/optimize.md /path/to/your-project/.claude/commands/
cp 02-memory/project-CLAUDE.md /path/to/your-project/CLAUDE.md
cp -r 03-skills/code-review-specialist ~/.claude/skills/

Understand what that means before you start. There is no plugin to update, no dependency to track and no compatibility check between the guide and your Claude Code version. Once optimize.md is in your .claude/commands and the code-review-specialist skill is in your ~/.claude/skills, they are files in your repository and your home directory, and when the guide changes them, nothing tells you. Teams that fork this guide, which the README counts as a sign of adoption, are specifically responding to that problem. If you copy the files, budget for reviewing them by hand, and put them under the same review process as your own code.

The first-run download is the step that breaks on a corporate network

The guide carries an installation note about Claude Code itself, and the detail that matters operationally is the last sentence. Starting in v2.1.113, Claude Code ships as a native per-platform binary for macOS, Linux and Windows, and npm install -g @anthropic-ai/claude-code still works because the native binary comes down as an optional dependency on first use. As of v2.1.116 those downloads come from https://downloads.claude.ai/claude-code-releases, and corporate proxies must allowlist that host. So the fifteen minute path has an unstated network prerequisite, and it is not a documentation detail: if the host is blocked, the first use of the tool fails at download time, and the guide contains no mirror, no offline install and no troubleshooting entry for it. A reader working from a locked-down corporate network will read the clone command, run it successfully, copy the templates, type /optimize, and only then discover that nothing works.

The release tags stopped in June while the README claims a September sync

The newest three tags are v2.1.160 dated 2026-06-02, v2.1.138 and v2.1.131 both dated 2026-05-09. The README, by contrast, states that the guide is actively maintained and synced with every Claude Code release, giving v2.1.283, September 2026, as the latest. The main branch was pushed on 2026-09-26, so the branch is current and the tags are not. The gap matters for anyone trying to cite a specific state of the material. If you clone a tag, you get a guide written against whatever Claude Code looked like in June, with no marker telling you which version that was. The version numbers in the tag names belong to the guide's own scheme, not to Claude Code's, so nothing in the tag itself lets you line it up against v2.1.283. Read from the branch, and use CHANGELOG.md to see what moved, rather than treating a tag as a compatibility marker.

The self-assessment runs inside the tool it is teaching you about

There are two ways to find your level: take the self-assessment quiz linked from the learning roadmap, or run /self-assessment inside Claude Code, which then returns a personalised path. The per-module check is similar, with /lesson-quiz [topic] run after each module to point at what you missed. This is a good design, because the feedback comes from the same runtime you are being taught, and the quiz commands are themselves slash commands, so you are practising the thing being explained. It also creates a circular entry requirement. Someone who has not yet installed or launched Claude Code cannot use either the in-tool assessment or the topic quiz, and the guide's own diagnosis of the problem starts from someone who has already installed it and run a few prompts. For a complete beginner the only route in is the static level table, and that is the table whose time estimates do not reconcile.

Five translations multiply the maintenance surface without adding a sync signal

The README is published in English, Vietnamese, Chinese, Ukrainian and Japanese, each with its own README in a vi, zh, uk or ja directory at the repository root. The module material, the numbered folders, the scripts and the docs directory exist once, in English. So a translated README is a translation of the map, and the destinations it points at are the same English files. That is an efficient arrangement and it also means drift is invisible from the outside: nothing in the repository listing shows when a translation was last updated relative to the English original, and a reader following the Chinese or Japanese path is reading localized prose that may describe a table, a time estimate or an installation note that has since changed. If you are working in one of those languages, check the English README and CHANGELOG.md for anything the localized copy asserts about versions, since the English file is the one that gets corrected when a fact moves.

The root directory is an author's desk, not a clean tutorial

Alongside the ten module folders the root carries CLAUDE.md, STYLE_GUIDE.md, clean-code-rules.md, claude_concepts_guide.md, resources.md, QUICK_REFERENCE.md, INDEX.md, CATALOG.md and LEARNING-ROADMAP.md, plus four documentation directories in docs, prompts, slides and resources, and a local-progress directory that looks like working state. There is also tooling config: .cspell.json, .markdownlint.json, .pre-commit-config.yaml, .gitissue.yml, a .claude directory and a scripts directory. A coverage.xml sits at the top level. None of that is wrong for a guide maintained in the open, and some of it is genuinely useful, since STYLE_GUIDE.md and clean-code-rules.md are the kind of thing teams ask for. The reader-facing cost is that a clone gives you lint configuration, a spellcheck dictionary and a coverage report alongside the templates. Clone with a shallow history if you only want the files, and be aware that some top-level documents describe the guide itself rather than Claude Code.

Editorial conclusion

claude-howto fits someone who already has Claude Code running and wants working files to copy rather than prose to read, especially if you like the Mermaid diagrams and the quiz loop. It does not fit someone whose network blocks the binary download host, since the guide assumes that host is reachable and documents no workaround, and it does not fit anyone who needs to know which Claude Code version a given snapshot targets, because the release tags stopped in June 2026 and the branch is the only current state. Before you follow the 15-minute path, check that downloads.claude.ai is allowed through your proxy, read the module from the branch rather than a tag, and expect to own every file you copy.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes