One thing I changed after fifty test snaps

During the build, I snapped a lot of test boards. My office whiteboard. Conference room boards at a coworking space I was a member of. Boards from demo sessions with friends. Fifty-plus captures before the first beta tester ever touched the app.

One thing kept appearing in the output that felt wrong every single time I saw it.

### The problem: present-tense summaries

The AI kept writing summaries in present tense.

"The team is discussing the Q3 strategy. The decision is to proceed with the enterprise tier. Marketing needs to align on the messaging."

This tense is wrong for a meeting that already happened. The board was from a meeting that finished. The decisions were already made. Writing about them in present tense implies the conversation is ongoing, or that the decisions haven't been committed to yet.

Past tense is the correct register for summarizing completed sessions. "The team discussed. The decision was made. Marketing was flagged as a dependency."

The difference sounds subtle. In practice, it changes how the summary reads and how confidently people treat the decisions in it. Present-tense summaries read as tentative. Past-tense summaries read as conclusions.

### Why the AI defaulted to present tense

Large language models default to the present tense when summarizing or describing, because that's the most common register in their training data. News articles describe events as they unfold. Explanatory content describes things as they are. Present tense is the statistical default.

For a meeting summary tool, that default is wrong. Meeting summaries are retrospective. Everything happened before the summary is written.

### The fix

One instruction added to the Summary Writer prompt:

"Write the summary in the past tense. The session has already occurred. The decisions have already been made. Describe what happened and what was decided, not what is happening or what is being decided."

The change took 30 seconds to make. The improvement was immediate.

Here's the same summary before and after:

Before: "The team is considering three pricing tiers. The enterprise option is being prioritized for Q3. Marketing is working on alignment."

After: "The team evaluated three pricing tiers and committed to prioritizing the enterprise option for Q3. Marketing alignment was flagged as a dependency before launch."

The after version is more confident, more actionable, and more accurate about the nature of the output.

### Why I'm writing about this

One prompt sentence changed the register of every summary BoardSnap produces. That's leverage.

I'm writing about this because most of the interesting improvements to AI products aren't about the model — they're about the instructions. The model is the same before and after. The instructions changed what the model does with the same capability.

The fifty-snap testing process was valuable not because it gave me bugs to fix, but because it gave me enough exposure to the output to notice patterns that weren't obvious from a few test cases. You need volume to see patterns. You need patterns to identify systemic problems. You need systemic fixes to improve all future output.

Find the thing that keeps feeling wrong. Then fix the thing, not the instance.