Run any AI agent skill at human pace. Observe friction, improve the skill as you go.
The consent model is socratic attestation: at every gate, the agent asks a judgment question about the specific risk at hand. The human's answer is the consent. No "Continue?" buttons. No rubber stamps.
Status: v0.1. One skill. Works in Claude Code today, portable to any agent that supports skills.
Agent skills are everywhere. Matt Pocock's collection crossed 176k stars. Superpowers, the Claude Code marketplace, community skill repos. People are writing increasingly powerful automations that plan, refactor, deploy, and review on their behalf.
But nobody is watching how these conversations actually go. Skills ship, agents run them, humans click through. Three things rot:
- The consent model. "Continue? [Y/n]" is the same gate whether the skill renames a variable or rewrites a database schema. The human stops reading.
- The conversational interface. Where does the agent over-explain? Where does the human get lost? Where does the flow break? Nobody observes this.
- The skill itself. Friction accumulates but never gets fixed. The skill runs the same way every time.
Slow Motion is user testing for your agent's conversational interface. Run the skill at human pace, observe the friction, and sharpen the skill based on what you find.
The consent model that comes out of this practice. Replace "Continue?" with a question that requires judgment.
Bad: "Ready to proceed with the migration?"
Good: "This rewrites 8,692 annotations. Anything in there we should snapshot first?"
The bad question is a rubber stamp. It can be answered without understanding what's about to happen. The good question targets the specific risk (data loss), invites domain knowledge (which annotations matter), and surfaces concerns the human might not have voiced unprompted.
The rules:
- The answer IS the consent. A substantive reply means the human is engaged and understood. No separate approval step.
- Target the risk, not the action. Don't describe what the agent is doing (they can see that). Ask about the thing that could go wrong.
- One line answers. If the question requires a paragraph, it's too broad.
- Correct misunderstandings before proceeding. If the human's answer reveals they missed something, stop. That's the gate working.
- Conversational, not pedantic. Colleague over coffee, not oral exam.
| Pattern | Why it fails |
|---|---|
| "Do you approve?" | Rubber stamp. No engagement required. |
| "Are you sure?" | Content-free. Sure about what? |
| "Ready to continue?" | "Go" in disguise. |
| "What module are we working on?" | Tests recall, not judgment. An exam question. |
| Any yes/no question | Binary answers require zero understanding. |
| Pattern | Why it works |
|---|---|
| "This deletes 3 files. Anything in there that isn't regenerable?" | Targets the risk (irreversible deletion). Invites domain knowledge. |
| "The skill skipped validation. Worth adding, or was that the right shortcut here?" | Surfaces a design choice. Lets the human override or confirm. |
| "We're about to hit the paid API at ~$540. Cheaper way to validate first?" | Specific cost. Invites creative alternatives. |
Slow Motion is a meta-skill: it wraps any other skill and runs it at human pace.
/slow-motion <skill-name> [args]
Two speeds:
| Speed | When | What changes |
|---|---|---|
| Observe | Default | Work proceeds. Note friction. Attest at existing gates. |
| Step-through | --step flag |
Pause before every file write, judgment call, phase transition. |
At each pause, the agent renders at three tiers (glance / click / ask). The human sees the one-line summary and the attestation question. They dig deeper only when something smells off.
After the run, the agent produces:
- An improvement log on disk (the audit trail)
- A concrete improvement diff for the skill (actual edits, not suggestions)
- Feedback capture for reusable behavioral rules
The skill gets better every time you run it through slow motion.
Signet is a separate tool for portable, in-document cryptographic attestations. The two compose at different layers:
| Layer | Tool | What it proves |
|---|---|---|
| Epistemic | Slow Motion | The human understood what they approved |
| Cryptographic | Signet | The approval is signed, portable, verifiable |
Slow Motion produces understanding. When that understanding results in a decision document, Signet stamps it. Together: you can't rubber-stamp because the gate demands judgment, and you can't fake the stamp because the cryptography is real.
They are separate tools because they solve separate problems. Not every slow-motion run produces a decision worth signing. Not every signet stamp needs a socratic gate. But when both apply, the combination is stronger than either alone.
npx skills add equanimitech/slow-motionOr clone and drop skills/slow-motion/SKILL.md into your agent's skills directory.
Slow Motion is an EquanimiTech instrument. It sits alongside:
- Signet -- portable cryptographic attestations for markdown
- Aperture -- attentional granularity for AI responses
- Lull & Learn -- agent-native spaced repetition
- Zenborg -- intention-cultivation garden
All instruments share one design commitment: the human stays the principal. The technology creates conditions for engagement, not shortcuts past it.
MIT