Outlier

Inspiration

Generative AI has dramatically raised the floor of creative work—but it can also pull everyone toward the same ceiling. While working with advertising concepts and YouTube scripts, I noticed that AI-generated outputs were often polished but interchangeable. They reused the same structures, emotional beats, visual metaphors, and “clever” twists. This is the deeper problem behind AI slop: not simply low-quality content, but creative convergence toward a statistically safe average. I built Outlier to answer a different question: instead of asking AI to generate one more idea, can AI first reveal its own defaults—and then help us deliberately escape them?

What it does

Outlier is a medium-agnostic anti-slop creative advisor for advertisements, scripts, campaigns, and other creative work. A user provides a brief or draft, along with optional “human grain”: personal experiences, uncomfortable truths, specific observations, constraints, or contradictions that should not be averaged away. Outlier then runs a six-stage process:

1. Brief Contract — identifies the creative objective, audience, medium, and constraints.
2. Baseline Ensemble — generates multiple plausible default responses.
3. Convergence Map — finds repeated structures, tropes, tones, and assumptions across those responses.
4. Break Strategies — proposes specific ways to escape the detected average.
5. Outlier Directions — creates several differentiated creative directions without producing randomness for its own sake.
6. Evidence and Evaluation — connects its claims to baseline evidence and evaluates each direction for divergence, coherence, specificity, intentionality, trope leakage, and “logo-swap” resistance.

The result is available as structured JSON for AI agents, readable terminal output, and an HTML report for human review. This makes Outlier useful both as a standalone advisor and as a reasoning layer inside a larger creative-agent workflow.

How I built it

I built the MVP as a TypeScript and Node.js command-line application. The system uses structured schemas throughout the pipeline so every stage can be validated and passed reliably to another agent. It generates a baseline ensemble, assigns evidence IDs to the generated samples, calculates the prevalence of recurring patterns, and uses that evidence to construct the Convergence Map. Outlier then generates break strategies and candidate directions before running a separate evaluation step. Coherence and specificity act as hard gates: an idea should not be considered successful merely because it is unusual. For the Build Week demo, the application calls Codex through the local codex exec workflow and reuses the user’s existing ChatGPT OAuth login. This allowed the demo application to work without embedding or distributing an OpenAI API key. Codex was also central to the development process. It helped turn the original “escape the AI average” concept into the six-stage architecture, implement the TypeScript pipeline, define the structured-output contracts, build the HTML report, create tests, and refine the product positioning.

Challenges

The biggest challenge was distinguishing meaningful divergence from random weirdness. If a system only rewards distance from the average, it can produce incoherent or generically surreal ideas. That simply creates a second kind of AI slop. Outlier therefore treats divergence, coherence, and specificity as separate dimensions. Another challenge was preventing the system from making unsupported claims about what is “common.” Rather than relying entirely on the model’s intuition, Outlier first generates several baseline responses and grounds its convergence analysis in observable repetitions across those samples. Evaluation was also difficult because the same model can favor its own stylistic habits. I addressed this by separating generation from evaluation, removing candidate labels during scoring, validating the returned score ranges, and requiring evidence for the final recommendations. Finally, the HTML visualization needed to communicate the difference between divergence and coherence without reducing creativity to a single score. The resulting map shows that the goal is not simply to move farther from the center—it is to find directions that are both distinctive and purposeful.

What I learned

The most important lesson was that human creativity is not created by adding arbitrary imperfections. What makes work feel human is often specificity: a real point of view, an unresolved tension, an inconvenient constraint, a memory, or an observation that would disappear during statistical averaging. I also learned that AI can be more valuable as a critic of its own defaults than as an endless generator. By making convergence visible before generation begins, we can use AI to widen the creative search space instead of narrowing it.

What’s next

The next step is to extend Outlier beyond text with multimodal analysis for images, video, audio, and presentation design. I also want to add team-specific convergence maps, historical style analysis, independent evaluator models, and integrations that allow other creative agents to call Outlier automatically before producing their final output. The long-term goal is not to make AI imitate humanity. It is to help people and AI preserve the choices, tensions, and specific points of view that make creative work worth remembering.

Built With

• openai
• typescript
• zod