Build your own workflow assistant – something that does one repetitive part of your job the same careful way every time. You describe the task in plain language; Claude interviews you, designs it with you, and builds it.
You do not need to know how to code.
No git, no plugin marketplace, no install step. Make an empty folder, paste one thing, and Claude walks you through the rest. About two minutes.
Before you start, you need Claude Code installed. If you don't have it yet, see claude.com/code. Everything below happens inside Claude Code – you'll be talking to Claude, not typing shell commands yourself.
Not sure how? Ask Claude: "Make a new folder called weekly-report in my Documents and open it." Naming it after the task you want to automate is a good habit.
Set up the agent-builder toolkit in my current folder, then walk me through starting it. Do these in order:
1. Run this command (macOS or Linux):
curl -sL https://github.com/anutron/agent-builder-plugin/archive/refs/heads/main.tar.gz | tar -xz --strip-components=1If I'm on Windows, run this instead:
curl.exe -sL https://github.com/anutron/agent-builder-plugin/archive/refs/heads/main.tar.gz -o main.tar.gz; tar -xzf main.tar.gz --strip-components=1; del main.tar.gz2. Confirm the folder now contains
README.md,CLAUDE.mdand a hidden.claudefolder, all at the top level rather than inside a sub-folder. If they ended up in a sub-folder, move everything up – including the hidden files – and delete the empty wrapper.3. Then tell me to type
/reload-skills, and explain in one line why it's needed. Wait for me to do it – you can't run it yourself.4. Once I confirm skills were added, tell me to type
/create-agentto begin.
That's the whole setup. Claude runs the download, checks it landed properly, and hands you the two commands in order – you'll be building a few minutes later.
Why those last two are yours to type: Claude can't run its own slash commands.
/reload-skills is needed because Claude Code looks for skills when a session
starts, so anything downloaded mid-session is invisible until you tell it to look
again. If your version doesn't have that command, closing and reopening the folder
does the same job.
/create-agent isn't recognized, in order of likelihood:
/reload-skillsdidn't run, or didn't report adding skills. Run it again, or restart the session.- The download landed in a sub-folder. Ask Claude: "Does
.claude/skills/create-agent/SKILL.mdexist here?" If not, have it look foragent-builder-plugin-mainand move the contents up – including the hidden.claudefolder, which is the part that matters and the part a plain move skips. - Still stuck? Ask in plain English: "Use the create-agent skill."
The download failed. Approve the network-access prompt if you see one. If
curl isn't found on Windows, use curl.exe as written above – plain curl means
something different in PowerShell.
/create-agent then walks you through a short interview and builds your first
workflow. Expect roughly 30–60 minutes, most of it you talking about how you
already do the task.
Tip: you can dictate rather than type. Type /voice tap, then press the
spacebar once to start speaking and again when you're done. The interview goes much
faster when you can just explain your work out loud – Claude will remind you when it
gets there.
On Windows? Everything works, but if you don't have git installed you'll be offered it during setup – it's a small download and it also gives Claude Code a better shell on Windows. Worth taking.
Agent-builder helps you create Claude Code workflows that:
- ✅ Gather information from several places at once
- ✅ Follow the same repeatable process every time
- ✅ Get better through use, rather than staying frozen at version one
- ✅ Include security and good practice by default
🧭 Guide's note – commentary built into this teaching tool, not Claude thinking out loud. It calls out real Claude Code features (Plan Mode, Auto mode, Skills, Subagents) as they come up, so you start recognizing them in your own projects.
⌨️ Your turn – a moment where Claude hands you the prompt to type instead of doing it silently. There are only a few, at the points where the technique is worth keeping: asking for a plan before anything gets built, getting a fresh-eyes review from a subagent, and making your first change yourself.
They're always optional – say "just do it" and Claude will. But prompting is the actual skill here, and you don't learn it by watching. The goal is that you leave able to do this without the walkthrough.
Running /create-agent guides you through:
- Setup – checks the toolkit is present and offers to set up saving (git) so your work is snapshotted as you go. You can decline; nothing else depends on it.
- Use case discovery – two paths: know what you want, or get help finding it.
- Process interview – understanding how you do the task today, one question at a time. Then Claude offers its own ideas before you commit to an approach.
- Data sources – what information the workflow needs, and whether it can reach it automatically or you'll paste it in for now. Pasting is fine.
- Architecture design – designed with you in Plan Mode, so you review and change the plan before a single file is written.
- Implementation – a working version one.
- Iteration planning – what version two should be.
- Final save – everything committed, with a history you can walk back through.
Everything downloads flat into your project folder, so these work immediately:
- Launches 5 agents simultaneously (~30–40s, vs 2–3 min one at a time)
- Duplication, simplification, conflicts, security, best practices, goal drift
- Shows findings; you choose what to track
- Run this after you've used your workflow, not just built it
- Captures what worked and what got in the way
- Updates the workflow's own files based on what was learned
- The habit this whole toolkit is built around: use it →
/improve-workflow→ use it again
- Reads improvements from
IMPROVEMENTS.md - Generates detailed commit messages
- Cleans up completed items
create-agent– the guided build, the/create-agententry pointworkflow-reviewer– orchestrates the 5 parallel review agentsworkflow-improver– the retrospective loop behind/improve-workflowsave-progress– git commits with real contextsecurity-checker– scans for secrets and credentialssoftware-best-practices– validates tests and linting, prevents goal drift
Stops Claude wandering away from what you actually asked for:
- Stop and re-evaluate every 20 minutes, or after 3 obstacles
GOAL.templatefor complex tasks, kept visible while working- Recovery questions when stuck, and pivot strategies when an approach isn't working
Work is snapshotted after each phase rather than once at the end, so there's always a known-good point to return to. Setting this up is offered at the very start and is entirely optional.
agent-builder-plugin/ # This repo – download it, it IS your .claude/ layout
├── .claude/
│ ├── settings.json # Project permissions, so you aren't prompted constantly
│ ├── commands/
│ │ ├── review-workflow.md
│ │ ├── save-workflow.md
│ │ └── improve-workflow.md
│ ├── skills/
│ │ ├── create-agent/ # /create-agent entry point
│ │ ├── workflow-reviewer/
│ │ ├── workflow-improver/
│ │ ├── save-progress/
│ │ ├── security-checker/
│ │ └── software-best-practices/
│ ├── agents/ # 5 parallel review agents
│ │ ├── review-duplication-simplification.md
│ │ ├── review-conflicts-setup.md
│ │ ├── review-security.md
│ │ ├── review-best-practices.md
│ │ └── review-goal-drift.md
│ ├── scripts/ # All shell lives here, not scattered in prose
│ │ ├── toolkit.sh # macOS / Linux
│ │ └── toolkit.ps1 # Windows
│ └── knowledge/
│ ├── workflow-patterns.md
│ ├── connectors.md
│ ├── component-decision-guide.md
│ ├── setup-command-guide.md
│ ├── security-guidelines.md
│ └── templates/ # File templates used by /create-agent
├── .gitignore
├── CLAUDE.md # How Claude should talk to you here – loaded automatically
├── CHANGELOG.md
├── LICENSE
└── README.md
CLAUDE.md is worth knowing about: Claude Code reads it at the start of every
session in this folder. It's why Claude skips the jargon with you, and it's where
your preferences get recorded as you make them – so a decision like "don't ask me
about git again" survives closing the window.
After /create-agent builds your V1, your project also has:
your-project/ # Same folder – agent-builder files above, plus:
├── project-plan/
│ ├── interview-notes.md
│ ├── project-design.md
│ ├── data-source-setup.md
│ └── IMPROVEMENTS.md
├── .claude/
│ ├── commands/
│ │ ├── setup.md # Generated by /create-agent
│ │ └── [your-workflow].md # Your workflow's entry point
│ ├── skills/ or agents/
│ │ └── [your-components] # Your workflow's own components
│ └── knowledge/
│ └── [your-references].md
├── CLAUDE.md
├── README.md # Replaces this one – it describes YOUR workflow
└── .env.example # Only if your workflow needs credentials
These are the shape of thing this is good at – a task you do on a schedule, that pulls from a couple of places, and where being thorough matters more than being clever.
- Input: last week's sales export, the staff schedule
- Gathers: sales by daypart, labour hours, notable comps and voids
- Output: a one-page summary in the format your GM actually reads
- Replaces: 90 minutes of spreadsheet wrangling every Monday
- Input: this week's invoices
- Gathers: line-item prices, compares against the last few weeks
- Output: what moved, by how much, and what it does to plate cost
- Replaces: noticing the cheese got expensive two months late
- Input: role, start date, location
- Gathers: the right handbook sections, checklists, training schedule
- Output: a personalized first-week packet, consistent every time
- Replaces: rebuilding it from the last person's copy and forgetting a step
- Input: the day's notes, open tickets, 86'd items
- Gathers: what's unresolved, what the next shift needs to know first
- Output: a short handoff the closing manager can skim in two minutes
- Replaces: a verbal handoff that half-survives
10 patterns common to workflows that actually get used:
- Phase-based workflows
- Parallel execution
- Specialized components
- Knowledge files
- State management
- Two-mode operation (guided and one-shot)
- Validation first
- Connecting to real data sources
- Research before action
- Permission configuration
See .claude/knowledge/workflow-patterns.md for details.
Once building and iterating on a workflow feels comfortable, there's more of Claude Code worth exploring:
- Global skills &
CLAUDE.md– conventions that apply across every project /fewer-permission-prompts– trim repeated permission prompts as you get more comfortable- Connectors – link Claude directly to Notion, Google Drive, Slack and the rest, so a workflow reads your real data instead of you pasting it. In Claude Desktop that's Settings → Connectors: pick a tool, sign in, done – nothing to install, and no API key stored anywhere.
/mcpdoes the same job in Claude Code /workflows– orchestrate bigger multi-step jobs once one workflow outgrows a single command- Aaron's public skill library –
github.com/anutron/ai, which includes/promotefor graduating a skill you like from one project to everywhere
Contributions welcome. See:
- Design patterns in
.claude/knowledge/workflow-patterns.md - Component decisions in
.claude/knowledge/component-decision-guide.md
MIT License