An AI Yoga Coach That Grades Your Pose From A Photo

If you’re trying to get better at anything in this world, you need constant feedback. This is definitely true for yoga: you might feel like you’re holding a pose right, but are you really? That’s where this private instructor and associated yoga challenge app comes in.

Pose Coach is a small AI agent that takes in a photo of you doing a pose, analyzes it, and gives you a score as well as advice. Photos are analyzed in memory and never stored. It lives on the Blocks agent network, so any person, app, or other AI agent can call it, and it runs around the clock on a small cloud server (Railway) with OpenAI’s vision model doing the visual analysis. The advantage of having this as a standalone agent is that you can do with it what you like. For instance you can see it in action here, and see how it’s extended into a challenge you can run with someone else: https://yogaposeoff.com/solo

Build Your Own

This has many applications, whether for wellness programs, fitness apps, yoga teachers offering async feedback between classes, and physical-therapy-adjacent coaching businesses. Basically it can help anyone whose product improves when a member gets instant, consistent feedback on body position.

Behind the scenes it works in three steps:

1. The user gives it a photo of their pose attempt — optionally with the pose name (“Tree Pose”) and a reference photo of the ideal form.
2. It analyzes the image with OpenAI’s vision model, guided by a coaching rubric that demands visible-evidence-only feedback and forbids made-up scores when no person is clearly in frame.
3. It returns a graded report card: score, strengths, corrections table, tips, and safety notes.

You can see here an example of the agent: https://app.blocks.ai/agents/pose_coach

You need three paid-or-free accounts to rebuild it: Blocks (agent network), OpenAI (vision analysis, paid per image), and Railway (hosting, a few dollars a month).

Starter Prompt

Copy this to an AI builder such as Claude Code:

Follow this guide to build and deploy an image-evaluation agent on the Blocks Network:
https://raw.githubusercontent.com/CagedEther/agentic-business-schooled/refs/heads/main/pose-coach-blocks-agent-skill.md

My goal: [e.g. a yoga pose coach / a deadlift form checker / a product photo QA agent]
Evaluation subject: [what the photos show and what "good" looks like]
Audience: [who reads the feedback and how strict it should be]
Deployment target: [Railway project name]
Output style: [score + corrections table + tips, or your own format]

I will provide my OpenAI API key and Blocks API key when you ask.
Document every secret you need, validate the agent card, deploy it,
and run one positive test (a real subject photo) and one negative test
(an unrelated image that should be politely declined) before reporting done.

You can access the skill file directly: reusable guide an AI builder can follow.

What to Try Next

Give the agent a memory: store each member’s scores over time and have it narrate improvement (“your hips are 10 points more level than last month”) — that turns a one-shot grader into a retention feature. Also you can extend the image recording capabilities to something else, like recognizing when your fridge is empty.