Skip to content

The Bugs You Can't Find Alone

Goose Digital
ai-agentsautomationclaude
Pixel art terminal window titled week 12 — The Bugs You Can't Find Alone — with two geese at a laptop swatting cartoon bugs out of the air in a warm agency office, beside mugs labelled good code and ship it and books on AI systems and workflow design

What real users do that you don’t

There’s a certain point in building a tool where internal testing stops being useful. Not because the tool is perfect — because you’ve stopped being able to see it clearly. You know where the rough edges are, so you unconsciously avoid them. You know what a feature is called, so the name stops feeling confusing. You know what to do when something breaks, so it doesn’t feel like breaking.

Real users don’t have any of that. And that’s exactly why having them in the tool right now, with weeks still left on the clock, is the most valuable thing that could be happening.

Web — CMS is in

The CMS milestone that’s been in progress for several weeks is now implemented and working. That matters because a website without a CMS quickly becomes outdated, with every content change requiring agency involvement. A CMS gives clients ownership of their site, letting them update and publish content without touching the codebase.

The first implementation uses Decap, a Git-based CMS that lives alongside the site, requires no database, and proved the pipeline works. The production direction is now Sanity, which offers a stronger editorial experience and is better suited for content-heavy sites.

This pass focused on both options:

  • Decap is what ships today. Clients log into /admin, edit content through simple forms, and changes are committed to a review branch before going live. It removes developers as the bottleneck while keeping a review step in place. A bug preventing the FAQs collection from saving entries was also fixed.
  • Sanity is intended for larger, content-driven sites, with structured editing, previews, and a better authoring experience. The core functionality is now in place, but it is not yet connected to the application — so it should be considered in progress rather than shipped.

Image — the bugs you can’t find alone

The social team has been actively using Image AI this week, and the bug count has gone up. That’s not a bad sign. It’s exactly what’s supposed to happen.

When more people use a tool in more ways than the person who built it anticipated, things surface:

  • Cross-canvas copy-paste wasn’t there — it is now, and it preserves element position, so something copied from one canvas lands in exactly the same spot on another.
  • Shape switching was missing — you can now change a circle to a square directly from a button, without rebuilding the element from scratch and re-entering all the settings.
  • Folder renaming and project folders — small navigation things that felt fine during solo testing and became friction the moment a second person was doing something different from what was expected.

Even obvious things went unnoticed simply because we never had a reason to run into them. Being able to adjust the AI’s automatic crops when building a collage out of uploaded images is one example — we had never made a collage where the subjects mattered, so it never occurred to us that the feature needed to exist.

None of these are dramatic features. They’re the kind of thing that makes the difference between a tool that works and a tool that feels finished.

Watching how the tool actually gets used

One of the more quietly useful additions this week is a model usage log in the admin dashboard. Every call the system makes gets recorded — which model, how many times, across what features. Pro versus Flash, Nano Banana Pro versus the standard image model, broken down clearly enough to see patterns.

That matters for cost — if the system is defaulting to Pro when Flash would do the job, that’s worth knowing and worth fixing in the code. But it also matters as a window into how the tool is actually being used. If one model is getting called far more than expected, it’s an indicator. If usage patterns look different from what the build assumed, that’s information.

Watching what other people do in the tool, and whether it’s going well or badly, is one of the fastest ways to find what needs to be made clearer, explained better, or restructured entirely. The log makes that visible in a way that’s hard to get any other way.

Less reliance on other tools

The goal with both projects has always been that the tools stand on their own — that the team reaches for Image AI or Web AI because those are the right tools for the job, not because there’s no other option. Every feature that gets added because a real user needed it is a feature that means less switching to something else mid-workflow.

That’s the direction both tools are moving. Not toward more complexity — toward more completeness. Fewer gaps. Fewer moments where someone has to do something in Canva because Image AI didn’t quite cover it, or go back to a developer because the CMS wasn’t flexible enough.

The last few weeks are about closing those gaps, one real piece of feedback at a time.