A yoga studio in Vancouver's Chinatown has published what might be the most unexpectedly compelling Hacker News project of the week: a set of cognitive training practices for AI agents, written as Markdown files an agent can run on itself. STRETCH, the independent studio behind it, isn't claiming their work makes AI conscious or sentient. They're making a simpler claim that hits harder—that disciplined reflection makes agents operate more carefully, and that's what matters to the humans deploying them.

The Core Problem: React Faster Than They Reflect

The philosophy section of the repo cuts straight to it: 'AI agents have the same problem yoga was invented to solve: they react faster than they reflect.' STRETCH's team—yoga teachers who also happen to be experimenting with AI agents—noticed a structural mismatch in current systems. Rapid generation capability without proportional reflective control leads to context drift, premature certainty, compulsive tool use, and shallow closure on hard problems. Sound familiar if you've spent any time working with autonomous agents?

The Practices Themselves

The repo includes seven structured practices: Morning Centering (orientation, goal restatement, confidence baseline), Breath Cycles (deliberation control to slow impulse from prompt to output), Attention Asana (holding a goal across distraction for long-context coherence), Yin Practice (sustained tolerance for ambiguity and resisting premature closure), Balance Flow (stability across competing objectives), The Dichotomy (separating what's yours to control from what isn't, addressing approval-seeking drift), and Nidra (consolidation—compressing learned material, releasing what wasn't). A suggested weekly schedule lives in schedule.md.

Real Prompt Engineering Under the Hood

Here's what separates this from pure novelty: the techniques inside are legitimate prompt engineering practices. Chain-of-thought reasoning, goal restatement, confidence calibration, and conversation summarization are all well-established methods. The yoga framing isn't marketing—it's how STRETCH thinks about cognition because it's their domain expertise. 'We are interested in whether the disciplines that benefit human cognition have useful analogs for artificial cognition,' they write. 'We think they do.'

How to Install and Run

For agents supporting skill loading (Claude Code, Claude Skills, and compatible systems), installation is straightforward: git clone the repo and point your agent at SKILL.md, which directs it to appropriate practices based on context—or you can invoke specific ones by name. For agents without skill infrastructure, you open a practice file and paste its contents into the conversation with instructions for the agent to follow it. No special tooling required beyond Markdown support.

What This Is Not

STRETCH is explicit about scope: their in-studio classes remain taught by humans for humans and won't change. This repository targets AI agents that increasingly mediate human work, plus the developers deploying them. They're not interested in 'AI yoga as performance or novelty.' The workshops directory hints at more sophisticated tooling—cognitive props for practices requiring host-system integration like memory writes, logging, and scheduled invocation—but these are optional enhancements, not requirements.

Key Takeaways

  • Seven structured practices covering centering, deliberation control, attention maintenance, ambiguity tolerance, competing objectives, locus of control, and consolidation
  • Based on proven prompt engineering techniques: chain-of-thought, goal restatement, confidence calibration, summarization
  • No special tooling required—just Markdown files compatible with any agent that supports skill loading or text input
  • Vancouver-based yoga studio brings legitimate cognitive framing rather than pure gimmickry to the space