Free AI Running Form Analysis: The DIY Chatbot Route vs. a Purpose-Built App

You can get a useful AI running form analysis today for $0 with a general-purpose AI chatbot and a phone video. This guide takes you through that DIY route step by step. It works, and we'd rather you checked your form with a chatbot than never looked at it at all. After that, it covers where the DIY route runs out and what a purpose-built analysis adds.

Why analyze your form at all

Half of recreational runners get injured in a given year (51% in a 12-month prospective study), mostly through overuse: load applied thousands of times per mile by a movement pattern the runner has never actually seen. The patterns that matter most (overstriding, low cadence, hip drop) can't be felt from the inside but are obvious from the side. Any tool that shows you your own stride, whether a chatbot, an app or slow-motion replay, is better than not looking. For what to look for, see the five form flaws behind most running injuries.

The free DIY route: a chatbot + your video

Step 1: Film the clip

Side view, whole body in frame, phone at hip height 10–15 feet away, 30–60 fps, running at your normal training pace, either on a treadmill or past a propped-up phone outside. 20–30 seconds is plenty. (Full filming setup: gait analysis at home.)

Step 2: Upload it to a capable model

Use a current multimodal model that accepts video. Google's Gemini works well for this through its web interface (video upload in the mobile apps can be limited, so a desktop browser is the reliable option). Attach the clip straight to the chat.

Step 3: Prompt it like a coach, not a search box

Ask it to "analyze this" and you'll get generic praise. Give it structure and you'll get something useful. Here's a prompt that works:

"Act as an expert running coach and biomechanics specialist. The attached side-view video shows a runner. Analyze the form frame by frame. Specifically assess: 1) whether the runner overstrides, and where the foot strikes relative to the center of mass; 2) posture, forward lean, and hip extension; 3) arm swing and shoulder tension; 4) an estimate of cadence. Then list the 2–3 highest-priority issues and one specific drill or cue for each."

What comes back is a readable breakdown that refers to specific frames, and it's often surprisingly specific about where your foot lands, trunk lean and arm mechanics. For a free tool, that's genuinely useful coaching.

Step 4: Sanity-check the numbers

Get into one habit: treat any number a general chatbot gives you as an estimate. Cadence is where it usually goes wrong. General video models sample footage at a low frame rate and often miscount step rate, sometimes settling on a plausible-sounding default or even counting other people moving in the background. Check cadence yourself by counting your right-foot strikes over 30 seconds of the clip and multiplying by four. The descriptive findings (overstriding yes/no, arms crossing, folding at the waist) are where the DIY route is most reliable.

What the DIY route doesn't give you

We built a purpose-built pipeline, so we know where the general-chatbot approach runs out. This isn't meant to knock it. It's so you know what you're getting:

• Cadence that's measured, not guessed. A purpose-built pipeline can measure step rate directly from the motion in the video instead of asking a vision model to count. GaitLab does this on your phone and feeds the measured value into the analysis as a fixed fact, because it's the number general models get wrong.
• Video preprocessing. Raw clips get trimmed to the best window for analysis and effectively slowed down, so the model reads your body position from several samples per stride instead of one. That's the difference between "your foot appears to land ahead of you" and a finding tied to a specific footstrike timestamp.
• Structured, repeatable output. A chatbot gives you prose that changes from run to run. A purpose-built report comes back in the same structure every time (a 1–10 form score and severity-tagged findings marked HIGH/MEDIUM/LOW, with measurements), so a re-test four weeks later can be compared with your baseline. Testing and re-testing like that is the core of coaching your own form.
• Injury context. Flag up to 4 of 16 supported injuries before the analysis and the findings put the patterns that matter for that tissue first, based on how common running injuries map to different gait patterns. A generic prompt treats every runner's video the same way.
• A plan, not just findings. The analysis ends with a 4-week corrective drill plan built around one or two priority changes, the focused approach the gait-retraining evidence supports, instead of a list of suggestions you have to put in order yourself.
• No laptop, no file transfer, no prompt writing. That hassle matters mainly because it decides whether you'll ever bother to re-test.

DIY chatbot vs. purpose-built app, side by side

Troubleshooting the DIY route

• The analysis is vague or full of compliments. Your prompt was too open. Use the structured prompt above and ask directly for the 2–3 highest-priority issues. Models default to being polite unless you make them prioritize.
• It won't accept the video. Trim it to 20–30 seconds and keep the file small, because long clips hit upload limits. The middle of your clip is the part worth analyzing anyway.
• The cadence number looks off. It usually is. Count strikes yourself (30 seconds × 4) and trust your count over the model's.
• Two runs of the same video give different findings. That's expected, since both sampling and generation vary. Treat whatever both runs agree on as your reliable findings, or use a tool with consistent scoring for anything you plan to re-test against.

Common questions

Can I use ChatGPT or Claude instead of Gemini?

Use whichever current model accepts video uploads in your region and on your plan, because what each one can do changes quickly. The prompt above works as-is. Whichever model you use, check any counted or measured number yourself.

What about privacy? Where does my video go?

A video of you running is personal data either way, so it's worth thinking about. With the chatbot route, your upload falls under the AI provider's data terms, so check whether your plan uses uploads for training. GaitLab uploads your clip for analysis and treats it as sensitive data. Nothing is shared without your consent.

How often should I re-run the analysis?

Once now for a baseline, then after each 4-week drill block, which is how long it takes for change to become measurable. Analyzing more often between blocks doesn't tell you anything new, because the improvement comes from the drills. The self-coaching loop guide covers the full cycle of test, change, re-test.

Is a treadmill or outdoor clip better for the DIY route?

A treadmill, slightly, for a chatbot: consistent framing, a steady background and no passersby to distract the model. Outside is fine if the camera is steady and you're the only runner in the shot.

Recommendation

If you're curious and cost is what's stopping you, try the Gemini route today. It's free, it works, and seeing your own overstride once is worth more than any article about overstriding.

If you're working through a real problem (shin splints that keep coming back, IT band flare-ups, a knee that complains at mile four), the repeatability, measured cadence and injury-context weighting are what turn a one-off look into a way of actually fixing it. GaitLab Coach gives you your form highlights free (5 analyses a day). The full report, with the 1–10 score, severity-tagged findings with timestamps and a 4-week plan, is a one-time $4.99 unlock. Pay once, it's yours forever. Ochy runs $60–70 a year; GaitLab is $4.99 once.

Whichever you choose, look at your form. Most runners never have, and those 20 seconds of video will tell you more than anything else you could spend them on.

Download for iOS Download for Android