Subir K

UX Rewired founder. Human-centered, AI-UX workflow creator. Off the beaten path traveler.

[Subir K @UXSubir]

Doubling context length cuts AI writing quality roughly 40%.
More unstructured context actively hurts performance.
The teams winning aren't giving AI more.
They're giving it less, better.
Are you engineering context?
Or just dumping documents and hoping?

[Subir K @UXSubir]

If you haven't figured out the cool AI tricks yet...
An annotated version of a visual I shared in Q4. Based on my conv. since then.
The ones who've got workflows are in the minority. Results, even less so.
So, if you're feeling FOMO from your email subs, videos, podcasts. Don't ;)

[Subir K @UXSubir]

Everyone's looking for prompt engineers to fix bad AI output.
Teams seeing 62% fewer inconsistencies didn't hire anyone new.
They cleaned up design tokens.
LLM optimized brand guidelines.
Their research specific.
Same skills. Different target.
Not a talent problem after all.

[Subir K @UXSubir]

The model isn't trying to deceive you. It's doing exactly what it was trained to do.

You're not asking it to be more careful. You're changing the reward structure so that honesty scores higher than confidence.

Which of these are actually in your prompts right now?

[Subir K @UXSubir]

AI doesn't lie because it's broken. It lies because in training, "I don't know" scores zero. So does being wrong.

So it guesses.

OpenAI's own research: models hallucinate on at least 20% of rare facts no matter what. The math guarantees it.

[Subir K @UXSubir]

Four more that shift the math:
1. Explicit callout: "Flag uncertain claims with [UNCERTAIN]."
2. Self-audit: "Identify your 3 least confident claims. Rate the response HIGH / MEDIUM / LOW."
3. Confidence sections: Require labeled CONFIDENT / PROBABLE / SPECULATIVE blocks.

[Subir K @UXSubir]

Here's an annotated version of a visual I shared in Q4. Based on a few months of AI-workflow convos with Prod/UX/Mktg ppl.
Prompting? Great. Workflows? Ad-hoc. Training? What training?
And results, KPIs? Measured in 'productive' terms.
In case, you're feeling FOMO. Don't ;)

[Subir K @UXSubir]

GIGO, PIPO still applies in the AI era.
When your AI outputs keep going off-brand, you give it MORE context, retry prompts.

The real cultrpits are the 60-page brand PDF, Word docs.
That's not signal.
That's noise wearing a document's clothing.

Less + Structured v.s. More & Verbose FTW.

[Subir K @UXSubir]

AI accelerates the requirements-design-build journey. That's the actual unlock.

Lower friction. Faster validation. One artifact to share with stakeholders, not three.

As a UX strategist, I can't wait to see smoother launch cycles, with more time for user validation.

[Subir K @UXSubir]

PM writes a detailed PRD. Design creates. Engineering reviews it, nods, and builds something slightly different.

They skimmed. Or were confused. By a requirement buried in story 32, p. 11. Or an implied use case.

This isn't a comprehension problem. It's formatting and overload.

[Subir K @UXSubir]

The handoff becomes more fluid. Devs and Design teams iterate.

Instead of torturous refinement meetings, you get real-time collaboration. A common workspace. Business sees real productivity gains.

Lower friction, higher throughput.

Shared ownership.

[Subir K @UXSubir]

AI can write your outline, draft your copy, and ship a feature by afternoon.
What it can't do: decide what shouldn't exist.
That judgment doesn't seem to be trainable.
It comes from knowing your user better than your backlog.

BUT. Are you practicing it IRL?

[Subir K @UXSubir]

The leverage won't go to whoever picked the right agentic tooling.

It'll go to who's experimenting with steering and delegation.
Like the dot-com era opened up new roles for the people embraced it.

And those who didn't dismiss the smartphone as a gimmick. But designed for it.

[Subir K @UXSubir]

The higher up you go, the more you trade understanding for leverage.

A manager sees the work. A director sees summaries. A CEO sees status. Each layer requires better judgment with less visibility.

Every PM working with AI agents just got promoted. Most aren't ready for it.

[Subir K @UXSubir]

AWS Kiro had outages where AI launched code without human review.
A friend of mine does cursory reviews before approving his PRs.

The hardest nut: approving plans you didn't build, trusting agents you can't audit, hoping for outcomes that your job rests on.
The risks are real.