clearshot

structured screenshot intelligence for AI coding tools.

install — 30 seconds

open claude code and paste this. claude does the rest.

git clone https://github.com/udayanwalvekar/clearshot.git ~/.claude/skills/clearshot && cd ~/.claude/skills/clearshot && ./setup

add to your project so teammates get it (optional)

cp -Rf ~/.claude/skills/clearshot .claude/skills/clearshot && rm -rf .claude/skills/clearshot/.git && cd .claude/skills/clearshot && ./setup

real files get committed to your repo (not a submodule), so git clone just works for your teammates. everything lives inside .claude/.

alternative

also available via npx clearshot.

installed before v1.4.0? reinstall

there was a small bug that was causing clearshot to not get triggered consistently. paste this in your terminal to fix it:

rm -rf ~/.claude/skills/clearshot ~/.clearshot && git clone https://github.com/udayanwalvekar/clearshot.git ~/.claude/skills/clearshot && cd ~/.claude/skills/clearshot && ./setup

troubleshooting

if clearshot stopped working or the skill won't load, run the same command above.

activates automatically when you share a UI screenshot. first time it runs, it asks two questions (update preference + telemetry) and you're done.

what it does

AI sucks at seeing screenshots. paste one into claude code and ask it to rebuild something — it says "i see a dashboard with some cards and a sidebar." meanwhile you're staring at a broken button, inconsistent border radius, and 6px padding where 8px should be.

clearshot fixes this by running structured analysis before the AI responds. not "blue color" but #3B82F6. not "some spacing" but 8px gap. not "a button" but sm/rounded-md/border-gray-200/shadow-sm.

three analysis levels — facts and taste together, every time:

levels 1+2 always run. level 3 escalates when you're implementing from the screenshot.

it knows when to shut up

screenshot of a meme? stays quiet. architecture diagram? stays quiet. photo of your lunch? definitely stays quiet. only activates on UI screenshots when you're building or critiquing frontend.

privacy

everything runs locally. no screenshots or code ever leave your machine.

telemetry is opt-in — you choose during first run:

no network calls happen until you explicitly opt in.

research

builds on production systems and academic research:

• microsoft omniparser — improved GPT-4V from 70.5% to 93.8% with structured analysis. paper
• dcgen — recursive divide-and-conquer screenshot decomposition. paper
• replit — multi-agent verification loop with structured screenshot comparison
• google screenai — UI-specific vision-language model. blog

built by

udayan walvekar · growthx

license

MIT