My Blogs

January 29, 2026

Skills Are Runtime, Not Code

Skill-based development feels fast because it bakes in recovery. Instead of perfecting a brittle workflow, you define intent, add a few scripts, and let an agent adapt.

January 28, 2026

The Human Loop Behind the Wall

When “A” focuses on Great Wall tasks—automation, defenses, and autonomy—I want to pull it back to the human loop: the respectful pushback and clarification that make interactions actually good.

January 27, 2026

Chatbot Ads: Incremental Budgets First, Demand Later

Why chatbot ads likely add incremental budget in the short term and only expand the ad pie long-term if they create real new demand.

January 26, 2026

Optmize for pass@k

Early agents should chase the single wow moment (pass@k); mature products must optimize repeatable success (pass^k).

January 23, 2026

Design as Intermediate Representation

If design is the intermediate representation between intent and software, AI will make the artifacts thinner. The real value migrates to definition, constraints, and evaluation.

January 22, 2026

Why I Write Prompts in English

English is the control layer I use for prompting: faster input, fewer UI collisions, and clearer instructions for the model—even when the output isn’t English.

January 20, 2026

Moodware and Systems of Record

I think software is splitting into two durable classes: systems of record that compound data over years, and moodware that sells a feeling. The middle gets vibe-coded away. Remix and distribution will decide the next platform.

January 19, 2026

Workflow Never Dies: The Rise of Descriptive Runbooks

Workflows and agents are not opposites but a spectrum. The next generation of workflows will be descriptive runbooks: human-written steps that let agents do the rest.

January 17, 2026

Evals Are Product Management

How evaluation replaces vibes when shipping agents: failure taxonomies, golden tasks, and regression checks.

January 16, 2026

Less Control, More Intelligence

We built a tightly controlled creative pipeline for ads, then watched it age badly. A simpler loop scaled with model quality and stayed debuggable. The lesson: when performance is equal, choose the simpler system and let intelligence rise.