From “this feels generated” to a judgement the next output can use.
Puddle & Patch is my independent project exploring how design judgement can reduce AI slop. I use a small multiplayer game to examine how appealing generated assets can still miss the intended identity, task and experience.
The valuable data is the reason behind each judgement. A preference without its task, evidence and trade off gives the next iteration very little to work with.
Brief
Define the player task and visual direction. The shared clay world should feel coherent while making the next action clear.
Output version
Keep the exact screen, state and image hash. Record the model, prompt and references when that provenance is available.
Exact region
Name the component and mark its location. Another reviewer should find the same evidence without guessing which element is meant.
Visible evidence
Describe the shape, material or relationship actually visible. Separate it from behaviour that a screenshot cannot establish.
Consequence
Connect the visible choice to the player task or agreed identity. Explain its effect instead of simply saying it looks wrong.
Preferred direction
State what should improve and what should remain. Give the next iteration a reusable principle rather than an unrelated redesign.
Verification result
Record the comparison or test, its result and remaining uncertainty. An untested preference should not become a measured fact.
Research basis explicit evaluation criteria and repeated critique are described in NN/G’s design critique research ↗. Recording the reasons behind preferences and retaining disagreement connects to CoVal’s preference rubric research ↗. This visual game study applies those ideas experimentally, their findings do not establish results for this project.
The same mismatch appears across the game.
The city, garden and racing images are baseline outputs under review. They are not examples of a finished clay design system. The annotations show which decisions need to change and which claims need an interactive test.
Carry the clay identity through the whole screen.
The direction changes the relationships between character surfaces, world signage, sculpted lettering, destination actions and a navigation family drawn from the game. The local player gets a ground ring, while other characters vary in pose.
This generated concept makes the proposed direction visible. It is a raster art direction reference, not an implemented interface or evidence of a trained model. Functional typography and interactive states still need an editable UI pass.
Original city design
Revised clay direction
Material and identity
Broader matte character forms, sculpted name plaques and a shared material treatment connect the scene to its controls.
Actions in the world
Destination names sit beside visible Enter actions. Map, friends and passport become recognisable sculpted objects.
Critique the revision too
Display lettering still has serif features, background detail remains rich and small labels need native size review. Improvement is criterion specific, not an automatic pass.
Can human judgement make the next AI output more coherent?
The game is the test environment. The work is to identify where a generated result misses the intended experience, explain the evidence and turn that judgement into criteria another output can be checked against.
The clearest gap is in this menu. A tactile world, a plush looking bunny and a flat interface do not yet share the soft clay direction. A palette alone cannot make them one visual system.
M01The interface does not share the world's material language.Revise
Where Central panel and action surfaces.
Evidence The scene uses textured, softly lit three dimensional surfaces, while the cream panel and green actions use smooth, regular flat shapes. Their edges and lighting do not carry the material treatment of the surrounding world.
Why it matters The menu reads as a separate interface layer rather than an intentional extension of the clay world specified in the art direction.
Direction Give panels and action surfaces a shared matte material, softened edge profile and restrained light direction. Keep text crisp.
Next check Compare the panel, primary button and scene side by side at the intended play size. Ask reviewers to identify the material cues, rather than simply choose the prettier option.
M02The heading introduces a different typographic voice.Revise
Where Local baking menu heading.
Evidence The main heading uses a compact, bold serif treatment. Its sharp terminals and editorial rhythm differ from the rounded forms of the character and environment.
Why it matters The strongest text treatment does little to establish the soft, sculpted personality intended for this game. A serif is not inherently wrong or evidence of AI authorship, the question is whether this choice supports this brief.
Direction Explore a rounded, substantial display voice for short headings, paired with a highly readable interface face. Limit sculpted or embossed treatment to large display text.
Next check Compare two typographic directions within otherwise identical menus. Record brand fit preference separately from reading accuracy.
M03The bunny's fine fur conflicts with the matte clay brief.Revise
Where Bunny cheeks, ear edges and torso.
Evidence Fine surface strands and fluffy edge detail make the bunny read as plush or fur. The project brief specifies a sculpted matte clay finish.
Why it matters Character and environment can share a palette yet suggest different construction materials. A rounded menu alone would leave this mismatch unresolved.
Direction Retain the folded ear, cream cheeks, expression and sage scarf, while reducing fur detail and using broader sculpted surfaces.
Next check A revised character asset is still required. Compare it against the canonical identity reference and use it in more than one pose.
M04The dock changes the visual language again.Revise
Where Bottom navigation symbols.
Evidence The bottom dock uses tiny, thin outlined utility symbols. Their line based construction differs from the volumetric objects and character in the scene.
Why it matters The navigation adds another visual vocabulary instead of reinforcing recognisable objects from the game world.
Direction Explore simplified clay inspired map, friend and passport symbols with a consistent silhouette and detail level. Keep labels and selected state indicators clear.
Next check Compare symbol recognition at native display size. Do not assume clay styling improves usability, and do not claim a touch target failure from this scaled crop.
M05The shared timer message competes with its own repetition.Revise
Where Subtitle, isolated clock and yellow status strip.
Evidence The subtitle says the shared timer is still running. A separate clock symbol and the yellow strip restate the same condition in additional rows.
Why it matters One important fact is distributed across several small treatments, leaving less attention for the return and leave decisions.
Direction Consolidate the multiplayer status into one conspicuous, readable message near the primary action. Verify comprehension before removing supporting cues.
Next check Ask players what continues while this menu is open. Record their answer and time to answer, a screenshot alone does not establish comprehension.
M06Keep the primary action more prominent than leaving.Keep
Where Return to kitchen and Leave activity actions.
Evidence Return to kitchen is the wider, darker action, Leave activity is smaller and lighter. The difference gives returning greater visual emphasis.
Why it matters The intended continuation action has a clear visual lead. Material styling should preserve this relationship instead of giving every control equal sculpted weight.
Direction Retain one dominant return action in the revised material study.
Next check Check first action identified and mistaken leave selections in an interactive prototype.
M07Keep the scarf to action colour connection.Keep
Where Sage scarf and green primary action.
Evidence The primary action repeats the muted green family used in the scarf, while cream appears in both character and interface.
Why it matters These repetitions already connect the interface to the character. The revision can strengthen material coherence without replacing the whole palette.
Direction Retain the cream and sage relationship while checking text contrast on the actual implementation.
Next check Review the palette across city, kitchen and garden. Measure source colours before making any formal contrast claim.
Game judgements and verification
| Judgement | Design decision | Next check |
|---|---|---|
| Make the menu belong to the bunny world | Carry broad matte surfaces, softened edges and consistent lighting into the panel, action buttons and icons. Preserve the darker return action so the material revision does not flatten the hierarchy. | Compare the revised menu with the original at the same size and content. Review material coherence separately from label readability. |
| Make the shared activity state truthful | Consolidate the running timer message and retain distinct pending and saved confirmations. Returning to the kitchen should be clear without implying that opening the menu paused everyone else. | Test opening the menu during a shared activity and delaying a reward save. Screens establish the intended wording, while runtime behaviour still needs verification. |
| Help the player recognise their own bunny | Retain the recognisable face, folded ear and scarf while varying poses. Use a clear local player marker, as explored by the ground ring in the revised city, without filling the plaza with more labels. | Check whether players can identify their bunny and cart during movement and overlap. The still images show the visual proposal rather than a measured recognition improvement. |
| Resolve the icon mismatch once across the game | Define one family of readable clay inspired icons and use it consistently across the dock and activity controls. Keep functional labels crisp and specify focus, pressed, disabled and selected states. | Inspect the icon family across all three activities at the actual play size. A matching material should not come at the cost of symbol recognition. |
| Check that the judgement improves a new scene | Record the criterion behind the change, such as controls sharing the agreed material while labels stay readable. Apply that rule to a new garden reward screen instead of copying the city composition. | Compare new outputs on the same task with model labels hidden. Track criterion level failures and regressions before claiming that guidance or a model change improved the result. |
| Give Puddle & Patch a recognisable brand voice | Define which shapes, display lettering and motifs belong to Puddle & Patch. Carry the folded ear, herb and carrot vocabulary into a restrained identity system instead of relying on generic rounded containers. | Compare the city, garden and kitchen with the logo covered. Ask whether their interface language still appears to belong to one game, without assuming recognition has already improved. |
| Let typography express play without slowing reading | Explore a rounded display face for short titles and keep a clear interface face for actions and status text. Establish consistent size, weight, spacing and line breaks so hierarchy survives when the same system reaches the garden and rally. | Review the title, primary action and supporting status at the intended display size. Compare brand fit separately from reading accuracy and avoid claiming that a typeface is inaccessible simply because it is serif. |
| Use colour to guide the task as well as the mood | Keep the palette relationship while defining separate roles for primary actions, supporting information, selected states and warnings. Use text or symbols alongside colour so readiness and saving do not depend on hue alone. | Measure contrast using the actual source foreground and background colours. Test whether the intended state remains understandable without colour, the scaled screenshots do not establish formal accessibility compliance. |
| Make the character feel grounded in the scene | Review feet, ground contact, shadow softness, camera scale and light direction as a set. Preserve identity while varying expression and pose, and keep the character surface broad enough to sit naturally beside clay props. | Inspect the character edges and contact shadows at full resolution. Review several placements together rather than accepting each asset in isolation, then check movement once animation exists. |
| Use open space to separate social play from scenery | Preserve the open plaza and keep activity entrances readable at the edges. Use spacing and restrained labels to separate nearby players, then place supporting systems where they do not compete with movement or destination choice. | Compare crowded and quiet player arrangements with the same camera. Check whether people can find a friend and an activity, a still composition alone does not establish wayfinding performance. |
| Make hierarchy respond to each activity | Define a hierarchy for each mode. Prioritise the shared task in the garden and the local cart, route and driving controls in the rally. Keep the broader navigation available without assuming it needs equal visual presence in every mode. | Use an interactive prototype to compare first identified action, missed controls and accidental navigation. Test under movement rather than inferring performance from an attractive static screen. |
| Preserve warmth through restraint and feedback | Keep expressive moments around discovery and successful play, while making waiting, errors and save confirmation direct and calm. Specify reduced motion and nonanimated alternatives when motion is added, rather than treating animation as a universal improvement. | Review the emotional tone across success, waiting and recovery states. Check motion comfort and comprehension in a playable prototype, the current artwork cannot demonstrate either. |
What the feedback should change in a future output.
These are testable expectations. They describe the proposed intervention and the evidence needed before claiming it worked.
Material consistency
Expected panels, buttons, icons and characters carry the same matte surface language.
Check blinded comparison on an unseen garden reward screen, with independent criterion ratings.
Brand recognition
Expected type, palette and silhouettes feel like one Puddle & Patch system across activities.
Check compare several scenes, preserve rejected alternatives and record reviewer reasons.
Clearer play decisions
Expected the local player, next action and saved progress state are easier to distinguish.
Check an interactive player task, with reading errors and mistaken actions recorded.
Generation guidance changes prompts and references. Training changes model weights through a compatible training method. Both need held out evaluation, a polished revision alone is not evidence of model learning.
AI made the options. I had to make the decisions.
Puddle & Patch is an AI assisted game concept where players meet in a bunny city, bake and grow herbs together. This study uses the generated artwork, typeset screens and design specifications to examine the role of human judgement.
Some outputs looked polished but made the next action harder to see. Others needed a different order or a clearer explanation. The examples below show what I noticed, the decision I made and what changed.
What I changed, and why.
I checked three things can the player see what to do, does the interface describe the real state, and does the sequence make sense? Each comparison keeps the evidence beside the decision.
Give the player room to find their way.
The city looked detailed, but the scenery competed with the player’s next action.
Before
After the judgement
- What did not work
- The generated drawing filled the plaza with detail, despite the brief asking for an open player area.
- My judgement
- Keep the bunny and activity entrances. Remove repeated buildings, paving and foliage that compete with them.
- What changed
- The revised sheet opens up the foreground and separates the entrances. I can point to the action without explaining the scenery first.
These are two saved storyboard versions. The later version also adds the third game. The comparison here is about composition, not a measured improvement in usability.
A menu should tell the truth about the game.
“Paused” made sense in an early solo draft. It became misleading when the game moved to shared play.
Before
After the judgement
- What did not work
- Keeping the old pause screen would tell a player the activity had stopped, while their teammates could still be playing.
- My judgement
- Treat this as a local menu. Explain what continues, then offer a clear way to return or leave.
- What changed
- The later screen is called “Local baking menu”. It explains that the shared timer keeps running and replaces “Resume” with “Return to kitchen”.
This is a correction for the multiplayer design, not proof that the model invented the original rule. The screens show the wording change. Runtime behaviour still needs a playtest.
Bake first. Decorate afterwards.
Two convincing images can still tell the wrong story when they are put in the wrong order.
Before
Before. Reading the generated atlas from left to right puts decorating before baking.
After the judgement
After. The selected shots follow the documented baking order.
- What did not work
- The atlas places the decorating shot before the oven shot. Used as a sequence, it would show frosting going on before baking.
- My judgement
- Review the task across frames. Keep the plain buns in the oven shot, then show frosting touching the baked pastry.
- What changed
- I put C05 before C04 in the presentation. The fix is the order of the existing shots, not a newly generated image.
The atlas is a collection of camera studies. This comparison demonstrates the assembly decision, not a claim that an animation or playable sequence was tested.
Work out the moment before polishing it.
The broad studies helped establish the world. The detailed storyboards then gave each scene a job where the player looks, what they do and what happens next.
I kept the early studies together, then worked through the city, kitchen and garden in more detail. These are AI generated pencil studies that I used to review the composition and flow.
Give the model a direction it can follow.
I chose a warm, tactile world with cream surfaces, sage accents and soft amber light. The same choices had to work across the bunny, kitchen, garden and interface.
For the character, I kept a short set of recognisable features cream cheeks, brown eyes, a pink nose, a sage scarf and one folded ear. A new pose still needed to look like the same bunny.
Cream
A quiet surface for labels and controls.
#FCF8EESage
The scarf and main interface actions.
#526747Soft green
Supporting panels and selected states.
#E6EBDFClay
Warm accents within the world.
#B86F5CAmber
Light, rewards and points of warmth.
#E7B866Ink
Readable text without hard black.
#343D2ECarry the composition into the finished scene.
The sketch set the camera and the action. The rendering pass added material, colour and light. I checked whether those additions still supported the original moment.
For close ups, the important detail was the contact a paw holding the tool, tea reaching the cup or frosting meeting the pastry. For the city, it was the open space between the entrances and players.
All seven observations, with their evidence.
“It looks nice” gives the next version very little to work with. A useful note names the element, points to the evidence, makes a judgement and explains why it matters.
Open a note to see the marked image, the reason and the next check. “Keep” records a choice worth preserving. “Revise” records a problem. A next check is still open until there is evidence that it passed.
Leave room for the player.
Revise+
Use the simplified sheet as the composition reference for the next render.
Keep the same character between poses.
Keep+
Compare every new pose against the same identity reference.
A polished render can still miss the material brief.
Revise+
Review a material revision beside the current render at the same scale.
Make the action physically clear.
Keep+
Use this frame as the contact reference for the pour animation.
Ready and present are different states.
Keep and test+
Verify the waiting and ready states in the playable build.
Finishing an activity is not the same as saving its reward.
Keep and test+
Test delayed saves and repeat checks against the same reward receipt.
Check the sequence, not just each frame.
Reorder+
Use the documented C05 → C04 order when assembling the final sequence.
How I would validate an evaluation+
The aim is to make a review repeatable and useful beyond one attractive image. Each judgement needs evidence, a consequence for the task and a check that could show the judgement was wrong.
- 01Define the task and criteria.
Agree what the output needs to communicate before reviewing it. Keep brand fit, readability and task behaviour as separate questions.
- 02Record the output and its source.
Keep the version, brief and image region with the observation. Record model and prompt details when they are known.
- 03Collect independent judgements.
Reviewers inspect the same brief privately before seeing other notes. They record the evidence and reason, rather than just a preference score.
- 04Preserve disagreement.
Group observations about the same issue without erasing different interpretations. Review uncertain groupings with a second analyst.
- 05Test whether the criterion transfers.
Apply the agreed criterion to an unseen scene. Compare equivalent tasks with model labels hidden and the order balanced.
- 06Report failures as well as improvements.
Record which criteria passed, which regressed and what remains uncertain. Keep guidance changes separate from any model training.
This project presents the process and worked design examples. Reviewer sessions, model training and comparative evaluation results have not yet been completed.
Detailed reviewer protocol+
What to check
Look at the whole experience, then record the areas with useful evidence. There is no quota and no requirement to agree with the proposed clay direction.
- Brand perception
- Does the game communicate a distinctive, warm social world? Identify which logo, character or interface choices support or weaken that impression.
- Aesthetic direction
- Do character, environment, lettering and controls share a deliberate material and lighting language? Describe the visible mismatch rather than simply calling it AI slop.
- Typography
- Review display personality, hierarchy, size, weight, line length and legibility. Which text treatment helps the player understand the next action?
- Colour
- Assess palette relationships, visual emphasis and foreground separation. Record measured contrast separately from an aesthetic preference.
- Accessibility
- Identify evidence about text readability and reliance on colour. Mark keyboard, screen reader, focus, motion and touch target questions for an interactive or source based check.
- Image integration
- Check cutout edges, scale, contact shadows, perspective and light direction. Does the bunny occupy the same world as the background and interface?
- Whitespace and density
- Is there enough room around players, entrances and controls? Name the competing elements and the task that needs space.
- Visual hierarchy
- What attracts attention first? Does the local player, current task and primary action have appropriate emphasis relative to scenery?
- Layout and alignment
- Assess grouping, margins, spacing rhythm and relationships between elements. Explain how a specific alignment affects scanning or comprehension.
- Character consistency
- Do the face, folded ear, scarf, proportions and material remain recognisable across poses and scenes? Separate intentional variation from identity drift.
- Icons and affordances
- Are the map, friends, activity controls and destination entrances recognisable? Explain what signals that an element is actionable.
- Interaction and feedback
- Review waiting, ready, active, local menu, pending save and saved states. Separate what the screenshot says from behaviour that has actually been tested.
- Task sequence and recovery
- Does mixing, baking and decorating follow a coherent sequence? What happens if a teammate leaves, saving is delayed or an action fails?
- Cross screen consistency
- Identify repeated patterns across city, kitchen, garden and rally. Decide whether one system rule can address several related observations.
- Emotional tone and restraint
- Which specific expressions, shapes or details support warmth and play? Which decorative choices compete with the task or make the experience generic?
Open the participant review workspace
Each contribution records the screen, exact region, visible evidence, judgement, consequence, suggested direction and verification needed. The final decision records what changes and why, preserving disagreements rather than replacing them with a single taste score.
Everyone receives the same screen and brief and writes privately before seeing other notes. Observations about the same element and claim are grouped without changing the original wording.
Each claim can be supported, opposed or marked uncertain. The final decision includes the evidence needed to revisit it. A second person checks the grouping.
Testing criteria on future output+
Use the agreed criteria on a new scene rather than repeating the same composition. Keep the task equivalent, hide model labels and balance the order of the outputs.
- 01Keep the evidence.
Save the brief, output version, exact region, judgement and rationale. Leave unknown model or prompt details marked as unknown.
- 02Test a new scene.
Apply the material, readability and task criteria to an unseen garden reward screen. Check the result without copying the kitchen layout.
- 03Compare the result.
Record passes, failures, ties and disagreements by task. Keep a visible improvement separate from behaviour that still needs testing.
- 04Update the guidance.
Change the design guidance only when the evidence supports it. Any model training or comparative evaluation remains a separate activity.
The choices come together in the game.
These baseline screens show where the generated assets and typeset interface still need direction. They are retained as evidence of the problems being reviewed. The annotations mark material, typography, icon and player identity mismatches rather than presenting the screens as a resolved design system.
Quality still needs someone to make the call.
AI makes it easier to produce options. The design work is deciding which choices serve the game, explaining the decision and checking that it survives the next iteration.
That is where I see human taste becoming more valuable setting a direction, noticing the mismatch, understanding the player’s task and giving feedback another person or a model can use.