# Issue 1: moving fast without inheriting the mess

URL: https://pixelactstudio.com/newsletter/2026-11

> Issue 1 of our newsletter: three checks before an AI-built prototype meets real users, a five-line weekly decision record, and how to pick what a pilot proves.

Issue 1. By Dev Talan. Published 2026-11-01.

Issue 1 of the Pixelact product engineering newsletter gives founders and product teams three practical notes: the checks to run before an AI-built prototype meets real users, a five-line decision record to write each week, and how to choose what a pilot should prove. It also covers EnvSift, a free Damn Labs tool for .env files.

Issue 1 of the Pixelact product engineering newsletter has three notes on moving fast without inheriting the mess: the checks to run before an AI-built prototype meets real users, why to write down one decision a week, and what a pilot should prove. Each note gives you something to do this week.

## Who is this issue for?

This issue of the Pixelact product engineering newsletter is for founders and product teams taking a product built fast, often with AI tools, to real users or a bigger build. The newsletter arrives once a month. Pixelact designs and builds MVPs, web apps, internal tools, and AI systems for funded startups, and these notes come from building our own products and the open source tools we publish under Damn Labs.

## What should you check before an AI-built prototype meets real users?

Check three things before the first real user signs in: that one account cannot read another account's data, that no secret key reaches the browser or the Git history, and that you can see an error and undo a bad release. Building the prototype with AI tools is a sensible way to test demand. These checks cover the move from demo data to other people's data.

- **Access control.** Create two test accounts. Sign in as the first, copy the URL or API request for one of its records, and replay it as the second. If the second account sees the record, the server is not checking who owns it. On Supabase, turn on Postgres row-level security and write a policy for every table the browser can reach. On your own API, check ownership on the server for every read and write. [Broken access control](https://top10.owasp.org/2025/A01_2025-Broken_Access_Control/) ranks first, as A01, in the OWASP Top 10:2025.
- **Secrets.** Next.js and Vite treat any variable prefixed `NEXT_PUBLIC_` or `VITE_` as public and build it into the browser bundle wherever client code uses it, so a Stripe secret key or a model provider key must never carry either prefix. Search the history too: `git log --all -p -S sk_live_` lists the commits that added or removed a Stripe live key, and a scanner such as Gitleaks searches past commits for many key formats at once. Rotate any key that was ever committed, because deleting the file leaves it in history.
- **Errors and rollback.** Send server and browser errors to a tracker such as Sentry, or at least to logs you can search. Write down the steps to redeploy the previous version and rehearse them once before launch. Back up the database and restore the backup to a scratch database, because rolling back a deploy does not undo a migration.

The [AI-built app risk checklist](https://pixelactstudio.com/resources/ai-built-app-risk-checklist) has 40 checks, from input validation to cost limits on AI APIs, and our guide to [the risks in AI-generated code](https://pixelactstudio.com/blog/ai-generated-code-risks) explains why these gaps show up so often in AI-built apps.

## Why write down one product decision a week?

A weekly decision record keeps the reasoning that code alone cannot show. Once a week, pick the one choice that would be hardest to reverse, such as the database, the authentication provider, how tenants are separated, or what data an AI feature may see, and write it down in five lines in the repository.

Code written with AI tools arrives quickly, and the reason behind it is the easiest part to lose. Six months later, a new engineer, a coding agent reading the repository, or an investor running technical diligence will ask why. Without a record, someone rebuilds the argument from memory or the team argues it again. Five lines take about ten minutes.

```markdown
# 0004: Use Postgres row-level security for tenant isolation

- Date: 2026-10-30
- Decision: Every tenant table has an org_id column and an RLS policy.
- Options rejected: A schema per tenant, or filtering in application code only.
- Why: One policy per table is easier to review than a filter in every query.
- Revisit when: A customer contract requires a separate database.
```

Number the files, and when you change course, add a new record that replaces the old one instead of editing it. Our guide to [architecture decision records](https://pixelactstudio.com/blog/architecture-decision-records) covers the fuller format and when to use it.

## What should a pilot prove before a bigger build?

A pilot should prove one thing: the assumption that would end the plan if it turned out to be wrong. Before any code, write down the question, the result that counts as a pass, and which results mean build, change direction, or stop.

**Matching the doubt to the pilot**

| What is in doubt | Pilot type | Example pass criterion |
| --- | --- | --- |
| Whether users understand and want the flow | Clickable prototype | Six of eight target users finish onboarding without help |
| Whether the hard technical part works | Technical proof | The supplier API returns order status in under two seconds for 95% of 500 test calls, on the plan tier you can afford |
| Whether automation saves real time | One automated workflow | A week of live invoices takes less staff time than the manual process, with every uncertain field sent to a person for review |

A pilot has grown too big when it tries to answer two questions or needs billing or a full permission model. Split it, and name the person who makes the call before it starts. A pilot that ends in a stop decision has still done its job, because it cost far less than the build it prevented.

Our [pilot sprints](https://pixelactstudio.com/services/pilot-sprints) follow this shape, and [prototype vs pilot vs MVP](https://pixelactstudio.com/blog/prototype-vs-pilot-vs-mvp) explains where each one fits.

## From the lab: EnvSift keeps .env files in step

EnvSift is a free, open source web app from Damn Labs, Pixelact's experimental and R&D branch. It compares, validates, merges, and converts `.env` files inside the browser tab, with no application server to receive them. Its `.env.example` Generator removes every value to make a template you can commit, which fits the secrets check above. ENV Compare lists the variables one environment has and another lacks, such as a key staging has and production is missing, with values hidden until you reveal them.

Version 0.2.0 shipped in July 2026 under the Apache License 2.0. Read [how EnvSift was built](https://pixelactstudio.com/work/envsift), or see every tool on the [Damn Labs page](https://pixelactstudio.com/labs).