claude-code-mastery: an obsolete-by-design guide whose hook scripts still earn their keep

A guide that declares itself obsolete on line one

The unusual starting point: this repository's own README headlines that Claude Code Mastery V2 is obsolete and directs readers to a newer Claude Code Mastery Starter Kit, dated 2026-02-14, plus a separate MDD package described as Manual-Driven Development with 21 modes. Previous guide versions V3, V4 and V5 live as pages on the author's site or GitHub Pages, each linked with UTM campaign parameters. So the reviewer's question shifts from is this guide good to what is still worth taking from a self-retired repository. The answer is the payload underneath the prose: GUIDE.md, a templates directory, five hook scripts, two skills and three command definitions. The guide portion covers global CLAUDE.md as a security gatekeeper, project scaffolding rules, MCP servers, the Context7 documentation integration, custom commands, single-purpose chats, and skills and hooks. For anyone evaluating it today, the obsolete banner is honest labelling, not a defect, and it frames the rest of this review.

The idea that survives: CLAUDE.md suggests, hooks enforce

The guide's core argument is the contrast between configuration layers. A rule written in CLAUDE.md is parsed by the model and weighed against other context, which means it may or may not be followed. A PreToolUse hook always runs, and its exit code decides the outcome: 0 allows the operation, 1 signals an error shown to the user, and 2 blocks the operation outright while feeding stderr back to the model. The README makes the point with a community quote saying a PreToolUse hook blocks secret-file access a few times per week, while Claude does not respect CLAUDE.md rules very rigorously. That is the durable insight here, and it generalizes beyond Claude Code: put policy in deterministic code where you can, and in prose instructions only where you must. The repository then applies the pattern across the tool's lifecycle, with PostToolUse hooks such as after-edit.sh running formatters, a Stop hook called end-of-turn.sh acting as a quality gate at turn end, and notify.sh raising desktop alerts. The exit-code table alone justifies keeping this page bookmarked.

Five scripts, two skills, three commands in the tree

The repository contents are small enough to enumerate. The hooks directory holds block-secrets.py, a PreToolUse gate that blocks access to .env files, block-dangerous-commands.sh, which blocks rm -rf and similar, end-of-turn.sh for quality gates, after-edit.sh for formatter runs, and notify.sh for notifications. The skills directory carries two SKILL.md packages, commit-messages for conventional commit generation and security-audit for vulnerability checks. The commands directory defines three slash commands: /new-project scaffolds a project, /security-check runs an audit, and /pre-commit runs quality gates. The templates directory supplies global-claude.md for ~/.claude/CLAUDE.md, project-claude.md for a repository-level starter, a settings.json hook configuration template, and a recommended .gitignore. It is a starter kit in the minimal sense, five files of enforcement plus scaffolding, and the scope is deliberately narrow: nothing here manages model selection, evaluation, or team-wide policy distribution.

Install is copy and paste, so audit what you copy

The installation flow is manual copying, documented step by step. You clone the repository, create the hooks directory with mkdir -p ~/.claude/hooks, copy the scripts in with cp hooks/* ~/.claude/hooks/, make them executable with chmod +x ~/.claude/hooks/*.sh, then copy templates/settings.json to ~/.claude/settings.json, with the README noting that an existing settings file needs a manual merge of the hooks section. Skills go to ~/.claude/skills with cp -r skills/* ~/.claude/skills/. Prerequisites are Claude Code itself, Python 3.8 or newer for the Python hook, and jq for JSON parsing in the shell hooks. Verification is starting claude and running /hooks and /skills to confirm loading. The README itself inserts the right warning, review and customize first, before copying the settings template. That warning deserves emphasis: this flow installs executable third-party scripts into a location a coding agent will invoke automatically, so reading block-dangerous-commands.sh before granting it authority is the reasonable price of admission.

The 39 percent claim and the one-task-one-chat rule

The guide's behavioural advice rests on cited research, and the citations are linked rather than merely invoked. For multi-turn conversations it references an arXiv paper, 2505.06120, for a 39 percent performance drop when topics are mixed in one conversation. For context rot it links Chroma's research on recall degrading as context grows, and for context pollution a post claiming that a 2 percent early misalignment can lead to a 40 percent failure rate. The derived rule is stated bluntly: one task, one chat. This review cannot verify those studies, and the numbers should be treated as the author's synthesis rather than settled fact, but the links are there for the reader to check, which is better practice than the usual unattributed assertion. The practical advice is cheap to follow regardless of the exact percentages: fresh sessions for unrelated tasks cost nothing and bound the blast radius of a polluted context.

MIT licence, quiet since May, and the author's funnel

Housekeeping facts for adopters. The repository is MIT licensed, so reusing and modifying the hooks and templates is unencumbered. The last push was 2026-05-07, four months before this review's reference date, there are no releases, and the README already flags the content as superseded, so treat it as a snapshot of Claude Code configuration practice from the first half of 2026. Note also what the repository funnels toward: the author's site with UTM-tagged links, the MDD package installed via npm install -g @thedecipherist/mdd, and the starter kit with its 16 slash commands, MongoDB wrapper, monitoring and layered security. That ecosystem is the maintained path, and this repository is the free, static distillation. The comparison worth making is with Anthropic's official Claude Code documentation, which stays current and explains the same mechanisms but ships no ready-made scripts; claude-code-mastery trades currency for convenience, handing you files you can copy in five minutes and audit in an hour.

Editorial conclusion

Claude Code users who want working PreToolUse scripts and settings templates to adapt should still clone this repository, because the exit-code enforcement model and the hook files are concrete and MIT licensed, but they must audit each script before copying it into ~/.claude and re-check behavior against the current Claude Code release, since the repository itself declares the guide obsolete and has been quiet since 2026-05-07. It is the wrong source for anyone seeking a maintained, current reference; that need is better served by Anthropic's official Claude Code documentation, or by the author's own successor projects, the starter kit and the @thedecipherist/mdd package. Verify before adopting: that the hook events and exit codes in templates/settings.json still match current Claude Code behavior, and that the successor links still resolve.

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. TheDecipherist/claude-code-mastery on GitHub