Ivan Likhuta
AI System Designer | GPT Agent Design | No-Code AI Prototyping | Behavioral Logic for LLMs
About:
I design GPT-based systems with structured behavior, token-level control, and production deployment in mind.
My focus areas include:
– Behavioral prompt architecture
– Instruction-resilient logic and agent consistency
– Memory scaffolding via Claude API + Redis
– No-code prototyping (Render, OpenWebUI, Framer, Make)
– Use-case-specific tools for writing, feedback, validation, and tone analysis
Behavior, in this context, means controllable, repeatable logic under ambiguity — not personality or style.
I build systems that follow instructions under pressure, resist invalid input, and can be tuned for tone, risk, or operational thresholds.
Background: 10+ years in field-based media and drone engineering; transitioned to AI in 2025.
Currently run Undrawn Labs — a solo R&D track for applied behavioral GPT systems.
I work with founders and product leads building GPT-powered tools under real-world constraints -logic, risk, cost, clarity.
Capabilities:
- Multi-layer prompt stacks with scoped behavior
- Role-conditional logic branching
- Instruction framing with persistent tone and structure
- Drift control across reworded or ambiguous prompts
- Boundary enforcement under high-pressure inputs
- Misalignment and overclaim filtering
- Context-resilient memory scaffolds
- Token flow control and cost optimization
- No-code prototyping
- Basic automation
- Tested in real workflows by copywriters, founders, and product leads
Design Principles:
- Behavior is a logic stack, not a style layer
- Constraints must be enforced, not suggested
- Systems should degrade predictably, not adapt silently
- Clarity is structural — not what’s said, but what holds
- Instruction drift is a design flaw, not user failure
Key Projects:
A set of diagnostic GPT agents I use to test prompt structure, behavioral logic, and communication clarity. Each agent is behaviorally fixed — not adaptive, not resettable. These are not products — they’re internal tools for high-friction validation.
Antagonist GPT
A one-shot behavioral agent built to test how well users maintain clarity and control under pressure.
It resists prompts, holds a strict tone, and doesn’t adapt. After 12 replies, its structure collapses - it reverts to generic ChatGPT and can’t recover.
What makes it unique: This isn’t interaction, it’s containment. The agent doesn’t respond, it holds. Its collapse is intentional, to test how prompt logic handles structural failure.
Use case: For testing prompt stability, resistance handling, and boundary-sensitive behavior design. Deployed in ChatGPT for testing - structure is platform-independent.
Deployed in ChatGPT for testing - structure is platform-independent.
Skeptic
A diagnostic agent designed to block soft logic, vague prompts, and unfounded claims.
It doesn’t assist or attack - it holds. The agent applies logical pressure without emotion. If a prompt lacks structure or clarity, it won’t engage. If it detects inflation or internal drift, it halts or redirects.
What makes it unique: Skeptic GPT doesn’t mirror tone or intent. It ignores emotional cues, rhetorical framing, and stylistic polish. Its only criteria: structural clarity and logical grounding. If those are missing — it doesn’t engage.
Use case: Used to test whether a statement holds under direct scrutiny. Helpful when surface phrasing seems solid, but the logic behind it may not stand up. Skeptic GPT forces a cold read: is the argument real, or just well-worded?
Deployed in ChatGPT for testing - structure is platform-independent.
PITCH BREAK
A diagnostic agent that stress-tests early-stage ideas through high-friction pitch simulation.
The agent simulates a pitch scenario, not a conversation. It takes the role of a disinterested investor and applies direct, friction-heavy questioning. You don’t get encouragement or guidance - only pressure. If your idea breaks, that’s the point.
What makes it unique: This isn’t critique for improvement - it’s designed collapse. The agent ignores tone, intent, and enthusiasm. It’s built to expose logical gaps and force structural clarity through resistance, not support.
Use case: Useful before presenting a concept, writing copy, or framing a new product. Reveals what sounds convincing, but doesn’t actually hold.
Deployed in ChatGPT for testing - structure is platform-independent.
Claim Check
A diagnostic agent for auditing claim strength in bios, product copy, and public-facing messaging.
It parses statements for clarity, specificity, and risk. Flags inflated phrases, vague projections, or misplaced tone. Doesn’t rewrite full text - only points out weak or high-risk claims and suggests tactical phrasing upgrades.
What makes it unique: This isn’t a writing tool. It doesn’t polish or inspire - it audits. The agent applies cold-read logic to identify where trust may drop: exaggeration, ambiguity, or tone misalignment. Suggestions are surgical, not stylistic.
Use case: Used before publishing messaging, especially in professional or investor-facing contexts. Helps ensure that the language holds up under skeptical reading.
Deployed in ChatGPT for testing - structure is platform-independent.
Work With Me:
Available for system design work on GPT-based tools - focused on behavior, control, and deployment.
Contact for project fit.
ivan.likhuta@gmail.com
Thank you!