01Principle 02 · Trust
Said "never stored". The code sends it with every scan.
- Plain words
- Says what a scan does
- Proof, one tap away
Live product, 1:44
SkillScout finds the AI tools that fit how you work.
200,000+ AI tools, 1,000+ new a day, and every list ranks the popular ones. I designed and built one that reads how you work.
87%Of your gaps get a pick that answers them, up from 70%
85%Of people who ran a scan opened a pick to read it
97%Of picks for non-technical people install in one click, up from 51%
// 01Your AI writes your profileCopy one prompt into your own AI. Paste its answer back. No account.
Who you are, in your AI's wordsThe gaps it found// 02It scans public sourcesEvery tool is scored against your gaps, not its stars. Nothing is saved.
Every tool scanned, down to 22Filtered by fit, need and trust// 03You get the few that fitPicks grouped by need. Each says why it fits, why to trust it, how to install.
Why it fits youWhy to trust itToo many tools, and nobody picks the few that fit you.
People who use ChatGPT or Claude every day keep hearing about skills, plugins and MCP servers. Most can't tell which ones fit their work, and every list ranks the popular ones. I hit it first, then my usability testers hit it in ways I hadn't planned for.
"I see 50 tools to select from, and I wanna throw away my laptop."
"It recommends me way too many design-to-code tools and I don't know their differences. I don't have the patience to go through each."
"What if my gap is in a different area which is not in my [Claude] history? … maybe I want to start looking at app development."
200,000+1,000+ new every day
111,810Agent skills
54,141AI tools listed
40,662MCP servers
Three rules I held to.
- Fit over popularityRank by how well a tool fills your gaps, not by its stars.
- Earn trust before the pasteNo account, nothing stored, privacy said where the worry happens.
- Explain every pickEach card says why it fits, why to trust it, how to install.
How it picks the few
- Rule 1Fit firstScored against your gaps. Popularity can move a score by a quarter, never more.
- Rule 2A floor of trustUnder 50 stars or 10 installs, it isn't shown. Official sources vouch for their own.
- Rule 3Your stack, not theirsTools built for someone else's setup sink.
- Rule 4No dead endsUp to 12 picks per need. Archived or dead links are dropped before you see them.
How a SkillScout scan works
- 12 days
- Solo
- Free tier only
- 60 s per source
- No outside AI model
Swipe the diagram sideways.
Your own AI writes a profile of how you work. It stays in your browser.
- Your own AI writes a profile of how you work. It stays in your browser.
- One request to the scan API. Redis checks the rate limit first.
- The scanner reads every public source at once, plus registry copies refreshed daily.
- Every tool is scored against your gaps: rules plus a small embedding model. No outside AI.
- Ranked picks stream back as each source answers. The profile is never stored.
45 explored in 4 days, 3 shipped.
Every option was explored between Oct 5 and Oct 8. The dashboard took eight versions; each one fixed what the last one got wrong, mostly how picks were ranked and grouped.
Dashboard · 8 versions
VER 1 OF 8
Scroll to flip, or pick a version.
Set asideExplored, then cut.
Click any screen for full size. Old versions re-rendered from their commits.
Four things that broke
Each one stopped a real person: a reader of the copy, me with my own paste, and a tester, twice. Each fix was a design call.
Show all four
02Access
My own paste failed. One strict box assumed you knew JSON.
- Saved anyway
- Says what it fixed
03Usability test
After saving, the next step was invisible. Run scan read as decoration.
- Your profile, shown first
- One next step
04Usability test
She needed help with something her AI had never heard of, and there was no way to say so.
- What her AI said, kept
- Add what it missed
Built in 12 days
99merged pull requests in 12 dayspeak: 52 on Oct 6Oct 5Oct 16
45design options explored in 4 days42 set aside3 shipped
8dashboard versions before the live oneOct 5Live, Oct 18
4breaks found by real people, all fixedreader · me · testerall fixed
A name, a mark, one screen at a time.
Brand
Round 1First mark
ShippedThe S, rebuilt in code
Favicon: ink on light, lime on dark
Palette
Ink#0B0B0B
Paper#E9E9E5
Lime#E6F77A
Mint#92E2A4
Amber#F3C75B
Coral#FF7B6B
Peach#FFB38A
Blue#C8D2E2
Fit on every pick: High mint, Medium amber, Low grey.
Type
200,000+Urbanist · headlines and big numbers
Picks by need, trust on every card.Inter · interface text
MCP · Skill · PluginSystem mono · tags and code
Lockups
SkillScout
SkillScout
The S mark with the Urbanist wordmark, lime on ink and ink on paper.
What I chose, and what it cost.
Four choices, four costs
No sign-up. Just paste.Paste what your AI wrote about you. Nothing to create, nothing to connect.CostNo sync across devices.Result100% saved a profile
Rules rank the picksEvery tool is scored against your gaps. A small built-in model re-ranks the top 200.CostPicks are only as smart as the rules.Result$0 in model fees per scan
Only tools you can trustEach pick needs a track record and a link that works.CostThis week's launches won't show up yet.Result97% install in one click, up from 51%
Anyone can use itNo install, no code, no setup. A designer, a marketer or a developer opens the site and uses the AI they already have.CostA website and a Claude skill to keep in step.ResultWorks with ChatGPT, Claude or Gemini
The people behind it.
What the testers said, and what I did
"Is it scanning after I've just created the profile or do I need to do something?"
The scan now starts from the profile popup. One clear next step.
"It doesn't look like any other vibe-coded product."
The look, and the story that leads into the app.
"I don't see a clear connection between the suggested apps/tools and the listed gaps."
Picks are grouped under the gap they answer, and each says why it fits.
"Profile step is super helpful and helps me reflect."
Your own AI writes the profile.
Feedback from testers in three recorded sessions and written notes.
What I'd take into the next version.
- The ranking is the product.
Every design decision sits on top of the scoring and the sources. Each scan pulls about 22,000 tools from many sources, and each source breaks in its own way. My own eval still gives the designer profile the weakest picks.
NextDeeper source connections, smarter scoring, and a cold scan well under 26.7 seconds.
- Testers find what I can't.
I designed for myself first. Two of the four breaks came from one tester's sessions, and the overload problem from another's notes.
NextFive designers testing it, recorded, before the next big change.
- Distribution is the hard part.
Building took 12 days. Getting people to it is harder, especially as a product designer. My launch post sent nobody to the site in its first day, and Google showed it 0 times in 28 days.
NextOne channel at a time, measured, starting where people already ask which AI tools to use.
