Most teams test their product from one side only. They read the code, or they click around the site. Bugs live in the gap between the two. This guide shows how to close that gap with two tools working at the same time: Claude Fable 5.2 reviewing your code and system from the inside, and Persqo's AI visitors using your live site or app from the outside, like real people with real goals.
We use this exact loop on Persqo itself. It is the most thorough way we know to check what you built.
The idea in one picture
| Side | Who does it | What it finds |
|---|---|---|
| Inside out | Fable 5.2 reads your code, config and data model | Logic bugs, security holes, money paths that charge twice, jobs that get stuck, settings that do not match each other |
| Outside in | Persqo AI visitors in a real browser | Buttons nobody can find, slow or broken pages, confusing copy, prices that differ between pages, forms that fail on a phone |
| The loop | Fable fixes, Persqo re-tests | Proof that the fix works for a real visitor, not just in a unit test |
Neither side is enough on its own. A clean code review does not tell you that your signup button is hidden on mobile. A smooth visitor session does not tell you that two webhooks arriving in the same second open two subscriptions.
Step 1: ask Fable 5.2 to criticize what you built
The single most useful thing you can ask a strong model is not "write more code". It is "tell me what is wrong with what I already have". Open your project in Claude Code, choose Fable 5.2, and ask for a review, not a feature.
Split the review into lanes. One giant "review everything" prompt produces a shallow list. Four focused reviews produce findings you can act on:
- Money: checkout, subscriptions, trials, coupons, taxes, refunds, who is allowed to change a plan.
- Security: who can see whose data, what an attacker can do with a public endpoint, which callbacks are not verified, where the server fetches URLs a user controls.
- Funnel and product: the path from landing page to first value to payment, broken links, untrue claims, emails that go out when they should not.
- Reliability: queues, retries, timeouts, stuck jobs, scheduled tasks, missing alerts, database indexes.
Here is a prompt you can paste for each lane. Change the lane name and the list of areas:
Do a read-only review of this codebase. Lane: MONEY. Scope: checkout, subscriptions, trials, coupons, tax, cancellation, refunds, and who in an account may change billing. Find real defects only. For each one give: - severity (P0 = money or security now, P1, P2) - file and line - a concrete scenario: inputs and state, then the wrong outcome - whether it is live in production, given our env flags - the smallest fix Trace every finding end to end in the code. Drop anything you cannot confirm. No style comments. At most 15 findings, ranked. Then list up to 5 opportunities separately.
Three rules make this work:
- Ask for a concrete scenario. "This could be a problem" is noise. "A customer pays, the return page and the webhook both run, two standing orders open, the customer is charged twice every month" is a finding.
- Ask it to drop what it cannot trace. A review that admits uncertainty is worth more than one that sounds sure about everything.
- Run lanes in parallel, and keep them independent. If you already ran a review, do not paste its results into the next one. A fresh, independent pass finds what the first one missed. When we ran a second round of reviews on Persqo with Fable 5.2, it found serious issues the earlier round had not, including a setting that hid the cancel button from paying customers while charges were already live.
Step 2: verify, then fix in small pieces
Treat every finding as a claim, not an order. Before you fix anything:
- Re-read the code path yourself, or ask Fable to re-verify the single finding against the current code.
- Check whether it is live. A bug behind a feature flag that is off can wait. A bug on the payment path cannot.
- Rank: money and security first, then anything a customer sees, then the rest.
Then fix in small, focused pull requests, one topic each, each with a test that fails before the fix and passes after it. A prompt that works well:
Fix finding #1 from the money review. First re-verify it against the current code; if it is wrong, say so and stop. Make the smallest change that fixes it. Add a test that reproduces the failure. Run the type check, the linter and the related test suites. Open one pull request for this finding only.
Step 3: connect Persqo to Claude with MCP
Now add the outside view. Persqo has an MCP server, so Claude can start tests, read findings and get fix prompts without you leaving the chat.
In Claude Code:
claude mcp add --transport http persqo https://persqo.com/api/mcp
Your browser opens, you sign in to Persqo, and the tools appear. If you prefer an API key, create one in Persqo under Settings, API keys, and add it as a header:
claude mcp add --transport http persqo https://persqo.com/api/mcp --header "Authorization: Bearer YOUR_KEY"
On claude.ai, open Settings, Connectors, Add custom connector, and use the same URL. Cursor and ChatGPT are covered on persqo.com/mcp.
Step 4: send AI visitors to your site, in parallel with the review
While Fable reviews the code, let Persqo test the product. Each test sends several AI visitors, each with a different profile, device and goal, through your live site in a real browser. They report what they did, where they got stuck, with screenshots, a replay and a fix prompt for each finding.
Paste this into the same Claude session:
Use the Persqo tools. Start a test on https://your-site.com with the goals a first-time visitor actually has: understand what we do, find the price, and sign up or contact us. When it finishes, list the findings by severity, and for each major one get the fix prompt.
For more on writing visitors that match your real buyers, read AI persona testing. For turning findings into prompts your editor can act on, read how to turn bug reports into Claude Code fix prompts.
Good goals are written in the visitor's words: "Find out if this works for a team of five", "See the total price with shipping before paying", "Book a call for next week". Persqo plans its own path from there, the way a real person would.
If your product is an app, point Persqo at the web app or at the landing page that sells it. If you have a staging site, use it: you get the same visitors without touching production.
Step 5: close the loop
This is where the two sides meet. Every Persqo finding becomes a task for Fable, and every Fable fix gets re-tested by a Persqo visitor:
- Persqo reports: "On a phone, the visitor could not find the pricing link in the menu."
- You ask Fable: "Here is the Persqo finding and its fix prompt. Find the cause in our code and fix it in one small pull request."
- After the deploy, you ask Claude: "Re-test that finding with Persqo." The same visitor walks the same path again.
- The finding closes only when the visitor succeeds.
The reverse works too. When Fable finds a bug in a flow, ask Persqo to walk that flow before and after the fix. A fix that passes unit tests but still fails for a visitor on a phone is not done.
What this looks like in practice
We ran this loop on Persqo, a live product with paying customers. In one round, four Fable 5.2 reviews in parallel returned about fifty findings across money, security, funnel and reliability. At the same time, Persqo's own visitors and a ten site pilot showed us problems no code review would catch: reports that contradicted themselves, a test that waited in a queue until it timed out, and a first report that took far longer than expected.
Every finding went into a small pull request with a test. Nothing was merged without passing the type check, the linter and the test suite. Then the visitors ran again.
A checklist you can copy
- Choose Fable 5.2 in Claude Code and run four independent read-only reviews: money, security, funnel, reliability.
- Ask for concrete scenarios and file and line, and tell it to drop what it cannot confirm.
- Verify each finding against the current code before fixing.
- Fix in small pull requests, one topic each, each with a test.
- Connect Persqo to Claude with MCP.
- Start a Persqo test with the real goals of a first-time visitor.
- Turn every Persqo finding into a fix task, and re-test every fix with Persqo.
- Schedule a weekly Persqo test and repeat the Fable review before every launch.
Start now
The outside half takes one minute: paste your URL into the free check and see what a first visitor experiences. Then connect Persqo to Claude from persqo.com/mcp and let Fable 5.2 and Persqo's visitors work on your product together, from both sides at once.
Questions people ask
Why combine a code review model with AI visitors?
They see different things. A model reading your code finds what is wrong inside: logic errors, missing checks, race conditions, money paths that can double charge. AI visitors in a real browser find what is wrong outside: the button nobody can find, the page that hangs on a phone, the price that differs between two pages. Most serious problems only show up from one side.
Do I need to be a developer to do this?
No, but it helps to have your project open in Claude Code or another agent that can read your repository. The Persqo side needs only a public URL. The prompts in this guide are written to be pasted as they are.
Does Persqo touch my live data or payments?
Persqo's visitors stop before payment and never enter real card details. On a site you have not verified, they only read public pages. Form tests, logins and deeper checks need proof that you own the site.
How do I connect Persqo to Claude?
In Claude Code, run: claude mcp add --transport http persqo https://persqo.com/api/mcp and sign in when the browser opens. On claude.ai, add a custom connector with the same URL in Settings, Connectors. Full steps are on persqo.com/mcp.
How often should I run this loop?
Run the full review before a launch and after any big change. Keep scheduled Persqo tests running every week, and ask Fable for a focused review of every pull request that touches money, login or data.