Vibeleaderboard
Index / tool

Skills for Real Engineers

github.com/mattpocock/skills
Visit github.com
Category
AI Agents
Rank
Pricing
Open Source
Platform
cli
Type
TOOL
Latest release
v1.2.3
Date

About

A collection of structured AI agent skills designed to improve software engineering practices and fix common failure modes in AI-powered development. Based on decades of engineering experience and focused on alignment, testing, and architecture.

What it does

A set of twenty-five instruction documents, written in plain markdown, that a coding assistant loads as named commands. Some fire only when a human types them and orchestrate a whole workflow; others hold reusable discipline the assistant can reach for on its own, and an orchestrator may call the reusable kind but never another orchestrator. There is no program to run: the mechanism is prose plus packaging, installed either as a managed read-only plugin that updates itself or as editable copies dropped into your own repository.

Why it's ranked here

The licensing is unambiguous: MIT, declared consistently in three places, so there is no doubt about forking or adapting it. It is released rather than dumped, with a real release process and a check that keeps the plugin manifest's version in step. Distribution runs through Claude Code's official marketplace, a channel with a review step. And the problem it addresses costs real hours, while the answer is a set of small documents you can read, adapt and delete individually rather than a process framework you must adopt whole.

What's good

The collection applies its own doctrine to itself. One document names the levers of writing for machine readers, keeping context small, revealing detail progressively, cutting filler words, avoiding negation, and the other documents visibly use them. They work as gates rather than advice: the debugging guide refuses to let the assistant guess at causes until it can name one command it has already run and show the output, and the testing guide forbids writing a test at a boundary the user has not confirmed. The included vocabulary file is a worked example of the artifact the documents ask you to produce.

Tradeoffs

Two of the top-level guidance files are byte-identical copies, exactly the duplication the writing guide argues against. Nothing here executes, so nothing verifies that a document still behaves as written: there is no test step, and whether a gate actually holds depends on the model reading it. The set is large and mostly human-invoked, so the human carries the index of what exists; the repository says this outright and ships a router document to compensate. A few instructions assume the author's own machine, pointing the assistant at a personal wiki under a home directory. And the code review flow is inert until a one-time setup step has run.

How to use it well

Pick one install route and commit to it: taking both leaves every document duplicated. The routes differ in ownership rather than content; the plugin updates itself but cannot be edited, while the copy route hands you files you own to adapt, which is the stated design intent. Run the setup document once per repository before anything else, because the review and triage flows read the configuration it writes, and write the project vocabulary file early since the testing and debugging guides consult it. Then start with two documents, not twenty-five: a grilling session before a change and the testing reference during it.

Technical notes+

The plugin's contents are enumerated in .claude-plugin/plugin.json, which lists 25 skill directories, 18 engineering and 7 productivity; AGENTS.md defines five buckets but only those two are promoted into the plugin. Each skill lives in its own directory with YAML frontmatter whose description string is the trigger the model matches against. .claude-plugin/marketplace.json makes the repository its own single-plugin marketplace as a fallback, and README.md documents the two mutually exclusive install routes with a warning that taking both duplicates every skill. For maintainers, scripts/link-skills.sh symlinks each skill directory into the local harness directories and bails out if the destination is itself a symlink back into the repo; it is marked dev-only. A one-time setup skill records the issue tracker and triage label vocabulary, and skills/engineering/code-review/SKILL.md reads that configuration to find the originating spec, pasting the smell baseline into its sub-agent prompt because the sub-agent has no other access to it. skills/engineering/diagnosing-bugs/SKILL.md blocks hypothesising until one already-run command and its output are named; skills/engineering/tdd/SKILL.md forbids a test at a seam the user has not confirmed. skills/productivity/writing-for-agents/SKILL.md names context load versus cognitive load, progressive disclosure, leading words and negation as a failure mode. package.json defines only changeset, version and check-plugin-version scripts, with no test script; AGENTS.md and CLAUDE.md are byte-identical in this snapshot.

Observed

License
MIT, declared in LICENSE, package.json and .claude-plugin/plugin.json
Contents
.claude-plugin/plugin.json lists 25 skill directories: 18 engineering, 7 productivity
Install surface
Two mutually exclusive routes: a read-only Claude Code plugin, or an installer that copies editable files into your repo
Tests
package.json defines no test script; the documents are prose with no executable verification
Executable code
The only shell script among the files read is scripts/link-skills.sh, marked dev-only and not a supported installer

Read from README.md, package.json, AGENTS.md, CLAUDE.md, CONTEXT.md, LICENSE, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, .agents/writing-docs.md, scripts/link-skills.sh, skills/engineering/tdd/SKILL.md, skills/engineering/code-review/SKILL.md, skills/engineering/codebase-design/SKILL.md, skills/engineering/diagnosing-bugs/SKILL.md, skills/productivity/writing-for-agents/SKILL.md.

What it can do

  • Fix misaligned AI-generated code

    Broken or misaligned code from AI developmentCorrected, properly aligned code

  • Reduce verbosity in AI-generated solutions

    Verbose AI-generated code or documentationConcise, optimized code or documentation

  • Debug and repair broken AI-generated code

    Non-functional code produced by AI agentsWorking, debugged code

  • Improve software architecture design

    Poor architecture patterns or designsWell-structured, scalable architecture recommendations

  • Generate structured testing practices

    Code requiring test coverageComprehensive test suites and testing strategies

  • Apply engineering best practices to AI development

    AI-powered development workflowsStructured, professional engineering practices

Tags

ai-agentsengineeringskillsclaudesoftware-developmenttddarchitecture

Tech Stack

Node.js

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.