Walnut Lab

From the Lab · Case study 01 · AI system build

Teaching AI what works,
instead of letting it guess.

Clickyness helps YouTube creators package videos: the title and thumbnail that decide whether anyone clicks. Under the hood is an AI system we built that studies which videos are massively overperforming right now, works out why, and turns those lessons into advice grounded in evidence. It runs in production, inside a product we own and operate.

Build sheet · CS-01Our own product
Product
Clickyness
Built for
YouTube creators
System
Retrieval-augmented generation
Parts
Scout · Translator · Coach
Channel calibration
≈ $0.04, one-time
Status
Live in production

Fig. 01 / The build at a glance

Full disclosure: Clickyness is ours. We built it, we operate it, and its AI bills land on our account every month. A case study means more when the builder has to live with the system.

S.01 / The problem

Generic AI writes the middle of the road. By design.

Ask a general tool like ChatGPT for a YouTube title and you get something that sounds fine and performs like an average. That’s not a flaw. These tools are trained on everything, so they produce the middle of everything.

But the difference between a video that gets ignored and one that takes off is knowledge that’s recent, specific, and constantly changing: what’s grabbing attention in this niche, this month. No AI model knows that out of the box. That information lives on YouTube itself, scattered across thousands of channels.

So we built a system to go get it.

views per uploadthe channel’s normal10x the channel’s normalflagged by the scoutrecent uploads · one channel

Fig. 02 / One upload, ten times the channel’s normal. That gap is the signal. Schematic.

S.02 / What we built

A scout, a translator, and a coach

Three parts, each with one job. Together they turn what’s working on YouTube right now into advice a creator can act on.

YOUTUBEthousands of channelsin the niche01THE SCOUTlearns each channel’snormal, flags the10x outliers02THE TRANSLATORextracts why eachoutlier workedthe source of truthTHE PLAYBOOKthe lessons, writtendown once per niche03THE COACHadvises from theplaybook + thechannel’s own datathe creator’s ownchannel data

Fig. 03 / The system, end to end. Copper marks the saved knowledge.

01

The scout

Watches the channels in a niche and learns what normal looks like for each one. When a video wildly outperforms its own channel’s baseline, say ten times its usual views, the scout flags it. Something about that packaging worked.

02

The translator

Studies each flagged video for why it worked: how the thumbnail was composed, how the title was built, what emotional angle it took. The lessons are written into a playbook the AI can read. Once per niche, then saved.

03

The coach

When a creator brings an idea, the AI doesn’t guess what a good thumbnail looks like. It advises from the saved playbook for that creator’s exact corner of YouTube, combined with the channel’s own data. Evidence, not vibes.

re·triev·al aug·ment·ed gen·er·a·tion /rag/ n. An AI system that looks things up before it answers, from a source of truth you control, instead of answering from memory. One of the highest-value systems a business can build. A plain-English guide: RAG: give your company a brain.

S.03 / Two decisions

Two decisions made it work

Neither one is about picking the right model. Both are about knowing where the value sits.

01

Learn the lessons. Never copy the homework.

We don’t feed competitor thumbnails into the AI to imitate. The system extracts the underlying patterns, then generates something original that applies them. Copying produces knockoffs. Pattern-learning produces work that’s both original and informed.

02

Do the expensive thinking once.

Analyzing an entire niche on every request would burn money. So the heavy analysis runs once and gets saved, and every request after that reads from the playbook. Calibrating the system to a new channel costs about four cents. Decisions like this are what let an AI system survive contact with a real budget.

cumulative analysis costrequests over timere-analyze on every requestanalyze once, save the lessons≈ $0.04 up front, then pennies per request

Fig. 04 / Cumulative analysis cost as requests grow. Schematic.

S.04 / The pattern

Now swap YouTube for whatever your business runs on

A distributor has quote histories and win-loss records. An ISP has service calls, coverage data, and churn patterns. An agency has every proposal it ever sent, and what happened next. Every business produces a steady stream of evidence about what actually works. Generic AI tools can’t see any of it.

The stream
For ClickynessThousands of YouTube channels, watched continuously.
For your businessQuotes, service calls, proposals, won and lost deals.
The signal
For ClickynessA video pulling 10x its channel’s normal views.
For your businessThe quote that closed in a day. The proposal that never got a reply.
The playbook
For ClickynessWhy the packaging worked, saved once per niche.
For your businessWhy you win and why you lose, written where AI can use it.
The output
For ClickynessTitles and thumbnails grounded in evidence.
For your businessDrafts, answers, and estimates grounded in your track record.

Fig. 05 / The same architecture, pointed at your business.

An AI system is only as good as what you connect it to.

Out of the box, AI gives everyone the same average answers. Connected to the data only you have, it becomes something a competitor can’t buy off the shelf. We build that connection.

Find out what this pattern looks like on your data.

Book a fit callTake the Fit Test instead

Fit call: 25 min · Founder-built · a few engagements at a time