SEVEIL · Browser · 2D card gameThree months of production learning
Built the discipline
Constitution · human acceptance · ownership · rollback
01 · Playable result · Aug 17–19, 2026 · Unity 6 · Solo direction + AI agents
Three days to a playableUnity 3D roguelite core.
I transferred an AI production system—not code—from a 2D web game into Unity 3D. Reliability came from bounded work, evidence and human acceptance—not blind trust in AI.
Final runtime scene + authored character composite
3Core prototype days
8:00Author playtest
58Enemies in targeted smoke
1UI pass hard-retired after review
02 · A portable production system
Not the code.
The management system.
SEVEIL and Silensee share no engine, medium or gameplay architecture. What survived was the way I turn intent into bounded, testable work.
SEVEIL · Browser · 2D card game→Engine-independent discipline
SILENSEE · Unity · 3D rogueliteThree days to prove portability
Reused the discipline
Bounded candidates · silent checks · visual gates · author playtest
DirectionOwner contractAI executionEvidenceHuman acceptance
03 · Reliability without pretending to be a programmer
I managed the codebase
through evidence.
I could not audit every C# line. I could still control production through contracts, plain-language reports, tests, rollback and human acceptance.
Human directionCanonicalOwner contractAI implementationSilent checksCross-chat reviewAuthor playtestAccept / reject
58Enemies materialized in a targeted silent smoke test
0Windows development build errors
V37Protected rollback baseline during V38 closure
Show the receipts — and their limits
SILENSEE_V38_SILENT_MAINTENANCE_PASS · materialized=58 · namedConfiguration=True · presentationSeparated=True · v37Protected=True · audioPaused=True · audioVolume=0.0
- 58 enemies is targeted smoke evidence, not a performance ceiling.
- The Windows development build recorded 0 errors and 488 existing warnings.
- V37 scene identity was protected by a recorded SHA-256 hash.
- Generic EditMode and PlayMode regression was not rerun during the V38 closure.
- This is a portfolio prototype, not a release candidate.
Three-day core versus post-core expansion
- Aug 17–19: V1–V38 established the playable core, playable build, rollback and eight-minute author playtest.
- Aug 20–21: V39–V49 slowed into controlled expansion—drone fix, blink, Q/E abilities, environment assets and recording polish.
- Case study, video editing and portfolio packaging are excluded from the development-time claim.
What cross-chat review meant
One AI executed a bounded task. Another checked its scope, Canonical, evidence and rollback story. I retained the final accept / reject decision.
Corrections are the proof
The build passed.
The experience still argued back.
Reliability did not mean pretending the first answer was right. It meant making wrong answers visible, reversible and cheap to reject.
Drone fire hit the player
Correction: collision ownership was isolated and retested instead of patched through presentation code.
The probe passed. The real horde was still wrong.
Correction: validation moved from an isolated enemy to the real spawned population chain.
The silence field looked like a basket
Correction: the visual owner changed while the gameplay contract stayed intact.
Impact VFX looked mechanical
Correction: gear-like, overbright shapes were replaced by restrained natural hits.
The build passed. Pressure failed at 5:00.
Correction: density, pursuit, enemy mix and upgrade variety moved ahead of polish.
A UI passed every test. I removed it anyway.
Correction: the rejected pass was hard-retired, protected and turned into a visual-gate rule.
04 · The playable promise
Move quietly. Trigger the crash.
Survive the horde.
One map, one character, one companion drone and one systemic rule: sound changes enemy state instead of flipping a scripted hostility switch.
Final runtime capture1 · Move quietly
Movement fills a visible noise meter. Standing still lets it recover.
Investigation → pursuit2 · Disturb the population
Footsteps, gunfire and world events enter the same sound-reaction system.
Late-loop pressure3 · Survive the horde
Automatic fire, upgrades and sound abilities turn stealth into controlled chaos.
05 · A sleeping population
Disturbed by sound.
A single enemy reacts locally. The horde emerges when alerted enemies wake neighbours and the reaction spreads.
This created a biomimetic crowd response without requiring a large AI model at runtime.
Autoplaying reaction model
Sleeping
Population exists before combat.
Local sound
A small pulse creates uncertainty.
Investigating ?
Nearby enemies enter a yellow search state.
Sonic boom
The expanding shock crosses the population.
Confirmed !
The horde converts to red pursuit together.
06 · Inclusive by design
A game about sound
that does not require hearing.
I wanted hearing-impaired players to read the same tactical information. I therefore designed and tested the prototype with audio muted.
Noise level, propagation rings, question marks, exclamation marks and enemy movement all carry meaning before sound is added.
Audio muted · system readable
?Investigating
!Confirmed threat
)))Propagation
07 · Visual evolution
The mechanics stayed private.
The visual language became legible.
This case study shows presentation decisions, not the underlying ability rules. Two early directions were rejected because they communicated motion and space too literally.
Only visual evolution is shown here. Ability parameters and interaction rules remain private.
08 · A map becomes a place
From one author sketch
to a navigable 3D night map.
The sketch already contained pressure, safe light, buildings, spawn clusters and escalation points. Production translated those relationships before polishing surfaces.
01 · Silhouette
Lock the playable footprint.
02 · Massing
Translate building relationships.
03 · Framing
Create readable boundaries.
04 · Ground
Add navigation texture and light zones.
05 · Occlusion
Shape reveal, shadow and threat.
How I reviewed the map — layer by layer
- Each stage isolated one visual question and ran its own silent EditMode check.
- Stage 02 caught three connected building shapes being interpreted as one object; the shapes were separated before the layer advanced.
- Only stages with explicit author approval are marked Author ✓. Later checks are evidence records, not retroactive approval claims.
09 · Prototype fidelity
Visual feedback without
a production-asset budget.
AI-generated, poly-budgeted prototype assets created a useful middle ground between crude placeholders and production-final art.
GenerateBudgetAuditCalibratePlaceTest
Prototype-ready is not production-final
- Exact triangle counts must come from the model audit, not visual estimation.
- Geometry alone does not prove performance; materials, textures, lights, shadows, draw calls and repetition still require profiling.
- Scale, pivot, collision, lighting and licensing remained explicit review gates.
A visible ten-minute-scale iteration
A chest became a readable reward.
01 · Archived runtime placeholder
Centered crop from the original V33 capture. Functional, but visually anonymous.
02 · AI candidate
A usable visual direction arrived in minutes.
03 · In-game common tier
Yellow communicates a standard upgrade.
04 · Premium candidate
Orange tests a rarer reward read in riskier spaces.
The common tier is shown in runtime. The orange image is an explicitly labeled color mockup, not a runtime capture.
10 · What the prototype proved
The method travelled.
Human judgment stayed in charge.
The same production discipline survived a new medium and engine.
Wrong answers became visible, reversible and cheap to reject.
AI executed. Evidence informed. I retained final acceptance.
What this case does not claim
This portfolio prototype does not claim release readiness, production-final assets, a performance ceiling, complete regression coverage or replacement of specialist engineering review.