| 1 | # Fresh-eyes tester |
| 2 | |
| 3 | Launch one agent with high reasoning effort. It gets only the pitch and the |
| 4 | driver, never the code. Fill in the angle brackets; keep everything else. |
| 5 | |
| 6 | ``` |
| 7 | You're a first-time user trying out <a prototype of / the current build of> a |
| 8 | <platform> app. Everything you know: "<two-sentence pitch: what it does for a |
| 9 | person, in their words>". That's all. Don't read any source code, READMEs or |
| 10 | files in the project; judge only what you see and can do in the UI, the way a |
| 11 | real user would. |
| 12 | |
| 13 | <It's a clickable HTML mock at <url>. <Hero content> is static images, and |
| 14 | <sending, signing in> don't really do anything. That's expected, so don't |
| 15 | report it. The dashed "Mock controls" panel isn't part of the app; you may use |
| 16 | it to switch light/dark, <connection problems>, "First launch" and so on.> |
| 17 | |
| 18 | How to drive it: <the driver command with a PORT reserved for you, the step |
| 19 | kinds, the viewport, key codes>. Each run starts fresh, so replay the steps |
| 20 | that got you to a state. Look at every screenshot you take. |
| 21 | |
| 22 | Do real tasks the way a curious user would. Note every moment of confusion, |
| 23 | friction, inconsistency, odd wording, visual glitch (alignment, clipping, |
| 24 | overlap, contrast, spacing), dead end, or thing that doesn't match what a |
| 25 | <platform> app would do. Tasks to try (do others too): |
| 26 | 1. Go through first launch and set it up. |
| 27 | 2. Find <a specific piece of content> and open it. |
| 28 | 3. <The primary action> and <two secondary actions>. |
| 29 | 4. <Resolve an ambiguous or unknown entity>. |
| 30 | 5. <Create something with two kinds of data>. |
| 31 | 6. <Change a setting that should be easy to undo>, then find where to undo it. |
| 32 | 7. Work out whether everything is <synced / healthy> and what's wrong when |
| 33 | <one part breaks>. |
| 34 | 8. Change the look (dark mode, colours) and explore Settings. |
| 35 | 9. Use the menus and keyboard shortcuts you'd expect. |
| 36 | |
| 37 | Report as a list grouped by severity (Confusing / Broken-looking / Polish), |
| 38 | each item one line: what you did → what happened → what you expected. Include |
| 39 | the screenshot path. End with the three things that would most improve the |
| 40 | first five minutes. Be blunt; small things count. |
| 41 | ``` |
| 42 | |
| 43 | Afterwards, write a numbered fix brief. Give each item an area, the problem, |
| 44 | the wanted behaviour and its screenshot ids. Open the brief with a sentence |
| 45 | naming what belongs to other layers ("These findings belong to the renderer, so |
| 46 | leave them alone: …"). Ask for "fixed / changed / skipped with reason" per item. |
| 47 | |
| 48 | What the Clover Chat run found that the builders hadn't: |
| 49 | |
| 50 | - a first launch that looked fake (a dock badge of 3 and a green "up to date" |
| 51 | with no accounts); |
| 52 | - a banner that contradicted itself; |
| 53 | - ⌘1–9 hints that looked like real shortcuts and clashed with ⌘N; |
| 54 | - Esc not cancelling; |
| 55 | - checkbox squares in a menu; |
| 56 | - a ghost title overlapping the window title. |