Skip to content
AFRICAN
TECH BROSay hello
FIND ME ONLINE Let’s talk
Back to the guide library
The Build LogCOOKIYFree guide

Validate your Claude Code app in a weekend (before you build the next feature)

You shipped, nobody bought. The Cookiy Skill drops into Claude Code, recruits 50 real strangers, runs the interviews, and gives you the answers your friends will never tell you. Here's the full loop.

By Akhil Boddu · AfricanTechBro

You shipped. Nobody bought. That's not bad luck.

Most builders shipping with Claude Code right now have the same week. Three weekends of vibe-coding, a clean codebase, a domain, a landing page, a launch tweet, and then silence. Four signups in the first week. Two are friends. One looks like a bot. One never opens the app twice.

The first instinct is to assume the product is wrong. Add a feature. Rewrite the hero copy. Run another launch tweet from a different angle. None of that fixes anything because none of it answers the question that actually matters: who is this for, and what would make them pay.

The validation step you skipped is the work. Not the building. The building is the easy part now. The hard part is finding 30 to 50 strangers who match your buyer, asking them the right questions, and reading the answers without flinching. That's what the Cookiy Skill does, end-to-end, from one Claude Code prompt.

What you'll have when you finish this guide: the Cookiy Skill installed in your Claude Code, your first study brief drafted, 30+ real interviews ordered, and a path from "I have findings" to "I shipped the fix."

Why asking 5 friends isn't validation

When your friend tries your app and says "yeah this is cool," they're not lying. They're managing the relationship. The same person on a Cookiy interview, getting paid, anonymous, talking to an AI moderator, will tell you the truth they wouldn't tell you over WhatsApp. That gap is the spiral of silence. People don't say what they actually think when there's social cost.

The other half of the problem is sample bias. The 5 people in your DMs already know you. They share your timezone, your context, your Twitter feed. They're not random strangers who'd find your landing page on Product Hunt at 3am with no context. The signal you need comes from people you've never met.

50 matched strangers, paid for their time, screened to your ICP, talking on video while an AI tracks tone and facial cues. That's the unlock.

What the Cookiy Skill does

Three things, all from one Claude Code prompt:

  1. Recruits real strangers matched to your ICP from a panel of 100M+ pre-verified respondents. Not your friends, not your followers, not Discord randoms.
  2. Runs the interviews with an AI moderator that asks your questions, follows up on interesting answers, reads facial expressions, and flags low-effort respondents automatically.
  3. Returns the findings as transcripts, sentiment heatmaps, and a written report with the top objections, the language your buyer uses, and the price they'd actually pay.

It's the missing primitive in the AI stack. There's an API for payments, comms, identity, compute. There hasn't been one for "what do 50 real humans actually think about this," priced for indie builders. Until now.

Cost reality: $0.50 per survey response, $9.90 per 15-minute video interview. A useful study costs $25 to $300. Compare that to Nielsen quoting $20K and 8 weeks for the same work.

The validation loop

This is the loop you're getting back. Most indie builders break it after step 2 and wonder why nothing's selling.

flowchart LR
    A[Build] --> B[Ship]
    B --> C{Signal?}
    C -->|Silence| D[Run Cookiy study]
    C -->|Conversion| K[Compound]
    D --> E[Read transcripts]
    E --> F[Identify the fix]
    F --> G[Ship the fix]
    G --> B
    K --> H[Next milestone]
    H --> A

Two rules. First, run a study every time you cross a milestone. Pre-launch, post-launch, post-pricing-change, post-feature-release. Second, never skip the loop because "I think I already know what they'll say." You don't. The whole point is the gap between what you think and what they say.

Step 1. Install the Cookiy Skill in Claude Code

The skill is open source. Two components, both required: the Cookiy Skill (workflow guide that tells Claude what to do) and the Cookiy MCP server (the backend that actually creates studies, recruits, and pulls reports).

Repo: github.com/cookiy-ai/user-research-skill

To install, paste this into Claude Code:

Install the Cookiy Skill from https://github.com/cookiy-ai/user-research-skill
and walk me through connecting the MCP server.

Claude reads the repo, installs both pieces, and prompts you for a Cookiy account. Sign up at cookiy.ai. They'll give you $30 in free credits to start (covers around 3 real interviews or 60 surveys).

[SCREENSHOT: Claude Code terminal mid-install, "Cookiy Skill installed, MCP server connected"]

Watch out: Both components have to be active. If you only install the Skill, Claude knows the workflow but can't actually create studies. If you only install the MCP, the backend works but Claude doesn't know how to use it. Confirm both are listed before you continue.

Step 2. Write the brief

The brief is the only thing in this whole loop that's hard to fake. The data is only as good as the screener. A loose screener gets you opinions from people who aren't your buyer, and those opinions tell you nothing.

Type this into Claude Code:

Connect to Cookiy and walk me through creating a study on:
"<your product, one sentence>"

ICP: <who you actually want to talk to. Be specific>
Sample: 30 interviews, English, global
Budget: $300 max

Claude generates a discussion guide. Read it before you launch. The thing to verify:

  • Tier 1 screener disqualifies anyone who doesn't fit your ICP. For a Claude Code app validation, that's "ships side projects monthly" + "currently pays for at least one developer tool." Hobbyists who ship one app a year and pay for nothing have a different price ceiling and will skew your data.
  • Volume floor. "Used the category in the past 30 days" is the bare minimum. Higher is better.
  • Geography filter if your ICP is regional. Cookiy panels default global, which is fine for most indie tools but breaks if you're positioning for a specific market.
  • Price question asked AFTER the pain question, never before. You want their pain in their words first, then the price calibrated to that pain.

Tip: Run Claude's AI-simulated interviews first. They're cheap, near-instant, and show you whether your discussion guide produces useful answers before you spend on real humans. If the simulated answers are vague or repetitive, your questions are wrong. Fix them. Then launch real recruitment.

Step 3. Run it

Once the brief looks tight:

Launch the study. Recruit 30 participants. Show me the cost preview before charging.

Claude shows you the preview. Confirm. Recruitment runs in the background. Most studies fill within 24 to 72 hours. You'll get notifications as interviews complete.

While it runs, do not pre-write the conclusions. Don't draft the "what I learned" thread. The whole point is being surprised. If you've already written the post in your head, you're going to read the data through that lens and miss what's actually there.

[SCREENSHOT: Cookiy dashboard, study running, 8/30 interviews complete, sentiment graph]

Step 4. Read the transcripts (don't just read the summary)

The summary report is useful. It's not the whole signal. The signal is in the verbatims. The exact words your buyer uses to describe the pain, the workaround they currently have, and what they'd switch from.

Read every transcript. Highlight three things:

  1. Repeated phrases. If 8 of 30 respondents independently say "I just want it to handle the ____ part," that's a feature spec written in their language.
  2. Ranked objections. What's the #1 reason they wouldn't pay? Not what's wrong with your idea, but what's wrong with the alternatives they currently use. That's where the gap lives.
  3. Pricing distribution. Don't average it. Look at the shape. If 40% would pay $0 and 30% would pay $20+, you have a free tier strategy and a paid tier strategy, not a pricing problem.

Watch out: A single low-volume respondent saying "I'd pay $5" doesn't tell you to lower your price. It tells you they're not your buyer. Anchor pricing to the volume they reported. High-volume users have a higher ceiling, full stop.

What good output looks like

Here's a real fragment from a noface.video study. 52 creators surveyed for under $25.

Finding What it says
59% use CapCut today The competitor is free, manual, and slow. Beat it on speed, not on features.
24% already glue ChatGPT to manual editing The exact gap noface.video closes. That's the wedge.
#1 frustration: "output looks generic and AI-made" The product roadmap priority. Beat this and you have a moat.
42% would pay $16+/mo The Pro plan can sit at $29 with no resistance.
32% want free Free tier mandatory, gates Pro.

This took 90 seconds to launch and 48 hours to come back. The same data from a traditional research agency would have cost $20K and taken 8 weeks. That's the asymmetry.

[SCREENSHOT: Cookiy report page, sentiment heatmap, top 3 objections, pricing distribution]

My noface.video case (the one this whole script is about)

Three weekends ago I built noface.video with friends. Faceless AI video tool. Code clean, idea solid, landing page sharp. We shipped. Almost nobody paid.

I was about to spend two months chasing creators in DMs for feedback. Slow path, low signal, social-cost-distorted answers. Instead I dropped the Cookiy Skill into Claude Code, wrote one brief, and let it run. 50 matched creators. Every objection. The exact pricing tiers they'd accept. The single feature most likely to make them switch.

The fix wasn't more features. It was tightening who I was selling to and changing the first line of the landing page. We shipped that the same week. The point isn't that one study saves a product. The point is that the loop replaces six months of guessing with two days of knowing.

Three mistakes most builders make

  1. Running a loose screener. "Has used the category" is too broad. Add a volume floor and a "currently pays for at least one tool in this space" filter. Hobbyists tell you nothing useful about pricing.
  2. Reading the summary instead of the verbatims. The summary smooths out the surprises. The verbatims are where the buyer's actual language is. Highlight the repeats.
  3. Treating it as a one-time study. A single Cookiy study isn't a strategy. Run one at every milestone, pre-launch, post-launch, post-pricing-change. The compounding signal is what makes the difference, not the one-off.

What's next

  • Run the study this week. Don't wait for "more product" first. The whole reason your launch is quiet is because you skipped this step. The fix is upstream of the next feature.
  • Read every verbatim. Twice. Once for the pain, once for the language.
  • Ship the fix and run another study at the next milestone. This is a loop, not a one-shot.

If you've shipped a Claude Code app and gotten crickets, the Cookiy Skill is the cheapest way to find out why. The skill is open source. Paste the repo into Claude Code and it walks you through the rest.

Sponsor note: This guide is part of the Labs sponsorship with Cookiy. Cookiy is the user research platform I actually use for noface.video. github.com/cookiy-ai/user-research-skill is the Skill repo. cookiy.ai signup gives you $30 in free credits, enough to run a real first study without paying anything yet.

Built something with Claude Code and got silence? DM me @africantechbro.labs on IG or TikTok with the link. If you run a Cookiy study off this guide, send me the findings. I'll feature the most useful ones.