AI Guides › Workbench
By Nigel Guy · 8 min read
Most people meet GitHub through a list post: ten repos, star counts, "bookmark this". You star the lot and install nothing. Or you paste a stranger's install line into your terminal without reading what it runs. Both feel productive. Neither is.
The rule: install one repo at a time, only after you have read what it will run on your machine, and only if it fixes a problem you had this week.
Everything below was checked against each repo's README or the Claude Code docs on 4 October 2026. Repos change weekly, so re-read the README before you run anything.
Claude Code's own docs are blunt: a plugin "can execute arbitrary code on your machine with your user privileges". Hooks and MCP servers run outside your permission rules and sandbox. So know two commands first:
claude plugin list shows what you have installed, with scope and status.claude mcp list (or /mcp inside a session) shows every MCP server and whether it connected.Inside a session, /plugin install name@marketplace opens a details pane with a Will install list, then asks for a scope: user (every project), project (shared with collaborators) or local (you, this repo only). Pick local while you are trying something out.
| Repo | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| kyrolabs/awesome-agents | Curated list of open-source agent frameworks and tools, grouped by category | Free | Seeing what already exists before you build | A list, not a vetting service; entries vary in quality and upkeep |
| ayghri/i-have-adhd | Skill and plugin that makes Claude lead with the next action and number its steps | Free (MIT) | Answers buried under preamble | Includes a SessionStart hook; read it before turning on always-on mode |
| eyaltoledano/claude-task-master | MCP server that turns a requirements doc into an ordered task list | Free; model usage billed by your provider | One large project you keep losing the thread of | Licence is MIT with Commons Clause, so you cannot resell it or host it as a service |
| YouMind-OpenLab/awesome-gpt-image-2 | Large open collection of prompts for OpenAI's GPT Image 2, with previews | Free to browse; image generation costs whatever your OpenAI plan or API charges | Starting image work from a working example | Run by a company (YouMind) that also hosts it on its own site |
| obra/superpowers | Skills that make Claude brainstorm, plan, test first, debug systematically and review before it finishes | Free (MIT) | Coding sessions that drift or stop half-done | Opinionated workflow; it adds context to your sessions |
The Claude-side tools need Claude Code, included in every paid Claude plan. Pro was around £18 a month including VAT at time of writing; checkout shows your exact price.
Nothing to install: use the table of contents (frameworks, evaluation, software development, research, memory, automation). Check it before you build your own agent; if three projects already do the job, you have saved the build. Run the buyer's check below on anything you pick.
In your shell:
claude plugin marketplace add ayghri/i-have-adhd
claude plugin install i-have-adhd@i-have-adhd
The first command adds the repo to your list of marketplaces, and the second pulls the plugin from it. Left alone, the install applies to your whole user account; append --scope local if you want it confined to a single project. Start a fresh session, type /, and the skill should appear in the menu. If you create an empty file at ~/.claude/.i-have-adhd-always, the hook keeps the style switched on for every session. Invoke it manually for a week before you commit to that.
claude mcp add taskmaster-ai -- npx -y task-master-ai
The -- passes everything after it to the server untouched; scope defaults to local. It needs one model provider's key, or none if routed through the Claude Code CLI, and works best from a requirements document at .taskmaster/docs/prd.txt. No doc yet? This prompt writes one:
You are a product manager writing a requirements document that an AI task planner will split into ordered build steps.
Project: [ONE_LINE_PROJECT_DESCRIPTION]
Who it is for: [TARGET_USER]
Must-have features: [FEATURE_LIST]
Out of scope: [THINGS_YOU_WILL_NOT_BUILD]
Tech constraints: [STACK_OR_HOSTING_LIMITS]
Steps:
1. If any input above is blank or vague, ask me up to five questions and stop. Do not invent features.
2. Write the document with these sections: Goal, Users, Features (each with acceptance criteria), Non-goals, Constraints, Open questions.
3. Write every acceptance criterion so someone could test it with a yes or no.
Format: plain text, under 800 words, no marketing language.
Before you answer, check: does every feature trace back to my inputs, and is anything here that I did not ask for? Remove it if so.
Fill in the five bracketed lines, then save the output as .taskmaster/docs/prd.txt.
Find a preview close to what you want and copy its prompt (gpt-image-2 is listed in OpenAI's API docs). Adapt it rather than pasting it unchanged:
You are an art director adapting an existing image prompt for a new brief.
Original prompt: [PASTED_PROMPT]
My subject: [WHAT_THE_IMAGE_SHOWS]
Where it will be used: [CHANNEL_AND_ASPECT_RATIO]
Text that must appear in the image, exactly: [EXACT_TEXT_OR_NONE]
Brand colours or style limits: [COLOURS_OR_RULES]
1. List which parts of the original are style (keep) and which are content (replace).
2. Rewrite the prompt for my subject, keeping the style elements.
3. Quote any required text exactly and say where it sits in the frame.
If any input is missing, ask before writing. Avoid named living artists, real people and trademarked characters.
Check before answering: is my exact text spelled correctly, and is the aspect ratio stated?
Inside a Claude Code session:
/plugin install superpowers@claude-plugins-official
The official marketplace is registered the first time you open an interactive session, so there is no marketplace step. Read Will install, including its context-cost estimate, and pick a scope. You get skills for brainstorming, planning, test-driven development, systematic debugging and code review.
| Repo | Why | Install |
|---|---|---|
| anthropics/skills | Anthropic's own example skills, plus the spec and a template | /plugin marketplace add anthropics/skills, then /plugin install document-skills@anthropic-agent-skills |
| anthropics/claude-plugins-official | The official marketplace's source; read a plugin here before installing it | Already registered |
| modelcontextprotocol/servers | Reference MCP servers: Fetch, Filesystem, Git, Memory, Time and others | Per server; the repo calls them educational examples, not production-ready |
| punkpeye/awesome-mcp-servers | Large community list of MCP servers | Per server; vet each one |
Run this before installing anything, in an assistant that can open the repo (or paste in the README, hook and config files).
You are a cautious security reviewer helping a non-specialist decide whether to install a GitHub project into Claude Code.
Repository: [REPO_URL]
How I plan to install it: [THE_INSTALL_COMMAND]
What I want it for: [YOUR_GOAL]
Work through these in order and cite the file or page for each answer:
1. In two sentences, what the project does.
2. Date of the latest commit and release, and whether issues get replies.
3. What runs on my machine: open hooks/hooks.json, .mcp.json, any bin/ folder and any install or postinstall script. Quote each command.
4. Whether anything downloads or executes further code at install or start-up.
5. What it can read, write or send: files, network, API keys.
6. Open or recent issues mentioning security, data loss, telemetry or broken installs.
7. The licence, and any limits on how I can use it.
If you cannot open a file, say so; do not guess its contents.
Output: a short table of the seven answers, then one verdict (Install, Install with care, or Skip) with the single main reason.
Before answering, check that every claim points to something you actually read.
Fill in the URL, install line and goal. For a cloned plugin, claude --plugin-dir <folder> plugin details <name> lists its components without starting a session.
Match a problem from the last seven days. Answers too long: i-have-adhd. Big build slipping: Task Master. Sessions ending half-done: Superpowers. Blank image canvas: the prompt collection. About to build an agent: the awesome list. None of those? Install nothing.
.claude/settings.json and turns the plugin on for everyone who works in that repo..mcp.json that you commit.