What this SOP does
This procedure gives your coding agent four upgrades. You never type a command. For each step you copy a plain-English instruction, paste it into your agent, and read what it reports back.
Graphify
A map of your project. Your agent reads less and finds things faster.
Ponytail
A short-code rule. Your agent writes only what is needed.
OmniRoute
A backup engine. When you hit a usage limit, work carries on with another model.
Agent Skills
A professional build process: plan first, build in small pieces, test, review, then ship.
Who this is for
- Founders who build with an AI coding agent and do not write code by hand
- Team members setting up a new machine or a new client project
What you need
- Claude Code, installed and signed in
- A project folder to work in (an existing project or an empty one)
- An internet connection
Do the steps in this order: Agent Skills, then Ponytail, then Graphify, then OmniRoute. OmniRoute needs the most setup, so it goes last.
Each dark block is an instruction for your agent. Press Copy, paste it into Claude Code, and press enter. Then compare what the agent reports with the "What you should see" box under it.
Open your agent and run a readiness check
The agent checks your machine and fixes gaps before anything is installed.
- Open Claude Code inside your project folder.
- Paste this instruction:
Before we install anything, check that this computer has Node.js, Python with the uv tool, and Git.
Tell me in plain English which ones are present and which are missing.
If anything is missing, explain what it is in one sentence and ask me before you install it.- Answer yes when the agent asks to install a missing tool.
A short plain-English list saying all three tools are present, or that the agent installed the missing ones.
Add Agent Skills
Gives your agent a professional build process. It covers writing a spec, planning, building, testing, reviewing and shipping.
- Paste this instruction into your agent:
Add the Agent Skills plugin by Addy Osmani. The only source to use is the GitHub repository addyosmani/agent-skills.
Install it so I can use the commands for specs, plans, builds, tests, reviews and shipping.
When it is done, list the commands I can now use in plain English.A list of nine commands: /spec, /plan, /build, /test, /constraints, /review, /webperf, /code-simplify and /ship.
You do not need all nine on every job. For a new feature, the useful run is spec, plan, build, test, review, ship.
Add Ponytail
Before writing anything, your agent asks whether the work can be skipped, reused or done in one line. It never cuts safety checks.
- Paste this instruction into your agent:
Add the Ponytail plugin. The only sources to trust are the GitHub repository DietrichGebert/ponytail and the npm package @dietrichgebert/ponytail. Other projects named ponytail exist, so do not use them.
Install it, turn it on, and tell me which level it is running at.Keep the two source names in the instruction. They stop your agent installing a lookalike by mistake.
A message saying Ponytail is installed and on, with its current level.
Useful things to ask Ponytail later
| Say this to your agent | What happens |
|---|---|
| "Run a Ponytail review of my latest changes." | You get a list of extras that can be deleted |
| "Run a Ponytail audit of the whole project." | The same check across everything, not just the latest changes |
| "Set Ponytail to strictest." | The agent writes the least code possible |
| "Turn Ponytail off." | Back to normal behaviour |
Add Graphify and map your project
Builds a map of how everything in your project connects, so your agent stops rereading the same files.
- Make sure Claude Code is open in the project you want mapped.
- Paste this instruction:
Install Graphify from the GitHub repository Graphify-Labs/graphify. The Python package name is graphifyy, spelled with two y's.
Connect it to Claude Code, then build the map for this project folder.
When it finishes, open the plain-English report it created and summarise the three most important things in my project.A new folder called graphify-out containing an interactive map, a plain-English report and the raw data. Your agent should also give you a short summary of your project.
Ask your agent to "keep the Graphify map up to date automatically whenever I save changes to Git". Then you never have to rebuild it.
The map marks each connection as read directly or guessed, so you can see how much to trust it. The project does not publish a token-saving percentage, so judge the saving on your own project.
Add OmniRoute as your backup engine
Runs on your own computer. When one AI model hits its limit or fails, work moves to the next one in your list.
- Paste this instruction into your agent:
Install OmniRoute from the GitHub repository diegosouzapw/OmniRoute (npm package: omniroute) and start it on this computer only.
Do not expose it to the internet.
Confirm it is running, then give me the address of its dashboard so I can open it.- Open the dashboard address your agent gives you. It should end in /dashboard.
- Add the AI providers you already use or want to use.
- Put your main model first, then cheaper models, then free ones. This order is your fallback chain.
OmniRoute sends your prompts and code to whichever provider handles the request. Free tiers often have different data terms than paid ones. Read the terms before routing client work through a free tier.
A cheaper or free model is also not the same as the one you started on. Fallback keeps you working, but quality can change.
Your agent confirms OmniRoute is running locally, and the dashboard opens in your browser.
Connect Claude Code to OmniRoute
Tells Claude Code to send its work through your backup engine.
- Paste this instruction into your agent:
Connect Claude Code to the OmniRoute gateway I just installed, following the OmniRoute guide for Claude Code.
Use the local address on port 20128 without adding /v1 on the end, and turn on model discovery so the gateway models show up in my model list.
Tell me exactly what you changed, and tell me if I need to restart Claude Code.- Restart Claude Code when your agent tells you to. The new settings load only at startup.
- Ask your agent to show the list of available models.
OmniRoute models appear in your model list after the restart.
OmniRoute also has a one-step launcher. You can ask your agent: "Show me how to start Claude Code through OmniRoute each day with one action."
Work the new way
From now on, each plugin does its job in the background while you talk to your agent normally.
| Stage of your work | Plugin doing the job |
|---|---|
| Starting on an existing project | Graphify maps it so your agent does not reread everything |
| Planning and building | Agent Skills supplies the spec, plan and build process |
| Writing the code | Ponytail keeps it minimal |
| Long working sessions | OmniRoute keeps you going when a limit hits |
Your first-week routine
- Day one: ask your agent to map the project with Graphify.
- Turn Ponytail on and leave it on.
- For each new feature, say: "Write a spec for this feature, then a plan, then build it one small piece at a time."
- On days you expect to hit limits, start Claude Code through OmniRoute.
- Before you ship, say: "Run a Ponytail review and a full code review on what we just built."
If your project has grown fast, ask for a Ponytail audit of the whole project. It finds extras across everything, not just the latest changes.
Confirm everything works
Ask the agent to check its own setup, then tick each item below.
- Paste this instruction into your agent:
Give me a plain-English status report on the four plugins we set up: Agent Skills, Ponytail, Graphify and OmniRoute.
For each one, tell me whether it is installed, whether it is working, and prove it with one quick test.
If anything is not working, tell me what is wrong and fix it, asking me first if the fix needs my approval.- Read the report and tick each item you can confirm:
All six ticked means you are set. Start your first feature by asking your agent for a spec.
Frequently asked questions
Do I need to run any commands myself?
No. You paste an instruction and your agent carries it out. You only make the decisions it asks about, such as approving an install or choosing providers in the OmniRoute dashboard.
What if my agent says it cannot do a step?
Some plugin installs are typed as commands that begin with a slash. If your agent says it cannot run one, ask it to show you the exact line, then paste that line into the Claude Code message box yourself. Everything else the agent handles.
Can I use these with other coding agents?
In most cases, yes. Graphify supports Claude Code and 15+ other agents including Cursor, Codex, Gemini CLI and GitHub Copilot. Agent Skills works in Cursor, Codex, Gemini CLI and Copilot. OmniRoute serves any tool that speaks the OpenAI format. Ponytail has install options for other agents in its repository.
Does Ponytail cut my token use in half?
No. The project's own test showed about 54 percent less code and about 22 percent fewer tokens. Use "cuts code by about half" if you talk about it publicly.
Is OmniRoute safe for client work?
Only if you have read each provider's data terms. Prompts and code go to whichever provider handles the request, and free tiers often have different terms than paid ones.
How many free tokens does OmniRoute give me?
The project claims up to about 1.6 billion free tokens a month by stacking free tiers. It re-audits that figure every two weeks and it depends on each provider's terms, so treat it as an estimate.
Troubleshooting
For every problem below, start by pasting the same instruction: "Something is not working. Find out why, explain it in plain English, and fix it. Ask me before you change anything important."
| What you notice | Likely cause | Tell your agent |
|---|---|---|
| Agent installs the wrong Ponytail | A lookalike project with the same name | "Remove it and reinstall only from DietrichGebert/ponytail." |
| Graphify not found | Package name typed with one y | "The package name is graphifyy with two y's. Reinstall it." |
| Agent Skills will not install over its first method | GitHub sign-in method not set up | "Try the HTTPS web address for addyosmani/agent-skills instead." |
| OmniRoute dashboard will not open | OmniRoute is not running | "Start OmniRoute and keep it running." |
| Claude Code ignores OmniRoute | Settings added after Claude Code started, or the wrong address | "Check the OmniRoute address has no /v1 on the end, then tell me when to restart." |
| Gateway models missing from the list | Model discovery is off | "Turn on gateway model discovery, then I will restart." |
| Answer quality dropped mid-session | OmniRoute moved to a cheaper or free model | "Show me which model answered, and help me reorder my fallback list." |
Source links and version
Graphifygithub.com/Graphify-Labs/graphify Ponytailgithub.com/DietrichGebert/ponytail OmniRouteomniroute.online Agent Skillsskills.addy.ie| Version | Date | Change |
|---|---|---|
| 1.1 | 8 October 2026 | Plain-English edition. Every step is a paste-in instruction for your coding agent. |
Want this set up around your own workflow?
Tell us what you are building and we will map these plugins to how you work.
Visit AI Savvy Founders