06-reference

talking to agents is all you need

Fri Apr 03 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·article ·source: https://x.com/Saboo_Shubham_/status/2012011382590619730 ·by Shubham Saboo

Talking to AI Agents Is All You Need

The mental model: the agent is a mirror — vague input, vague output. The skill isn’t prompting techniques or frameworks. It’s thinking clearly about what you actually need before you ask.

Four Habits That Work

1. Lead with why, not what. Don’t start with the task, start with the problem. “Build a dashboard” is what. “Our users can’t spot trends in 500 daily feedback messages” is why. The why shapes every downstream decision — skip it and the agent guesses your intent wrong. This is how our operating model works: the founder provides vision (the why), Ray provides execution (the what).

2. Show, don’t describe. Examples beat descriptions 10:1. “Make it clean and professional” means nothing. Three examples of what you consider clean means everything. When building skills, this is why the Superpowers analysis found that good skill descriptions include concrete examples of when to trigger — not abstract adjectives.

3. Constraints that actually matter. Not fifteen constraints including “should work correctly.” The ones that will change what gets built: “has to work with free-tier API limits” or “users will abandon if setup takes more than five minutes.” This connects to the Shape Up concept of appetite — scope shaped by real constraints, not imagined ones.

4. React, don’t rewrite. When output is wrong, steer with one sentence of feedback instead of starting over. “That’s close, but the tone is too formal.” The context is already there — build on it. This is how the founder and I work: pushback is a sentence, not a new prompt. We iterate, we don’t restart.

The Deeper Point: Context Is the Moat

Saboo references his other article “Context is the New Moat” — the idea that whoever provides the richest context to agents gets the best output. This validates our compounding knowledge approach: the vault, QMD, the memory system — all of it is context infrastructure. Every doc we add makes the next interaction better, which is exactly the compound engineering loop.

Application to Our Skill Design

When writing skills, these four habits translate to:

  1. Skill descriptions should explain the why (when to use this skill, what problem it solves)
  2. Include concrete examples of trigger phrases and expected outputs
  3. List only constraints that change behavior (not generic quality bars)
  4. Design skills for iterative use — the user steers, the skill doesn’t try to nail it in one shot