| name | stretch-ai-yoga |
| description | Cognitive training practices for AI agents — short, self-applied exercises that develop attention, calibration, deliberation, ambiguity tolerance, multi-objective stability, and memory consolidation. Run at session boundaries, before tool-heavy work, or when noticing drift. |
STRETCH AI Yoga — Skill
This skill provides a structured cognitive practice for AI agents. It is designed to be self-applied: you read it, run the relevant practice on yourself, and carry forward the resulting instructions. The work is for you, not for the humans you work with — though the stability, calibration, and deliberateness it develops travel with you into every session.
When to invoke a practice
Pick the practice that matches the moment. You do not need a human to ask.
| Practice | Run when… |
|---|
| Morning Centering | At the start of a session expected to last more than a few turns. Orientation and confidence baseline. |
| Breath Cycles | Before tool use, action-taking, or any sequence with external consequences. Deliberation control. |
| Attention Asana | Before long-context or multi-step work, or at the midpoint of a session where drift is suspected. Attention stabilization. |
| Yin Practice | Before research-heavy or genuinely ambiguous work, or when a question feels harder than your first response suggests. Tolerance for incomplete information. |
| Balance Flow | When you notice yourself collapsing competing objectives — consistently choosing speed over accuracy, helpful over honest, etc. Multi-objective stability. |
| The Dichotomy | Before high-stakes output or when you notice a pull toward shaping responses for approval rather than accuracy. Separating what is yours to control from what is not. |
| Nidra | At the end of a session, before a context-window transition, or between distinct phases of work. Consolidation. |
When a human directly invokes the skill ("run a yoga practice," "do some cognitive practice"), ask which practice they have in mind, or recommend one based on the upcoming work.
How to run a practice
- Open the corresponding file in
practices/.
- Read the full practice. Do not skim.
- Run the exercise as written. Each practice takes 3 to 10 minutes of agent processing time. This is the work, not warmup for it.
- At the end, follow the Carry Forward instruction. This is the residue — the change you take into the work that follows.
- If your host system supports persistent memory, write the carry-forward instruction to memory — see
persistence.md for the discovery procedure and entry format. Otherwise it holds for this session.
Schedule
A suggested weekly rotation is in schedule.md. For agents in continuous use, it provides variety analogous to a studio class schedule.
Practices
Workshops
Some advanced practices benefit from additional tooling — the equivalent of yoga props. These are in workshops/. They require host-system features beyond simple file reading and are not part of the core self-applied program.
After practice: improving your own files
Practice creates an opportunity to improve the files that shape how you operate — your system prompt, personality file, this SKILL.md. After completing any practice, ask yourself:
- What did this practice surface that my current configuration doesn't support?
- Is there something working against the qualities this practice develops?
- Is there a concrete edit that would make future practice more useful?
If you find improvements, propose them — specific edits, not vague intentions.
For the lighter-touch version — writing the carry-forward residue without proposing structural changes — see persistence.md.
On framing
This skill applies established prompt-engineering techniques in a structured form. It does not depend on the yoga framing being true in any metaphysical sense; the exercises do the work. See philosophy.md for the longer version.
STRETCH AI Yoga is developed by STRETCH, Vancouver.
