experiment

name	experiment
description	Chaotic domain: act to impose minimal constraints when cause-effect is invisible. Use when: production down with gibberish logs, completely novel problem, crisis, deliberate disruption, nothing makes sense. NOT for problems with any visible pattern — use /probe instead.
allowed-tools	AskUserQuestion, Read, Glob, Grep, Bash, EnterWorktree
argument-hint	<situation or crisis description>
cynefin-domain	chaotic
cynefin-verb	act
model	opus
context	main

Experiment

Act → sense → gate. Human IS the constraint. No autonomous runs.

Situation: $ARGUMENTS

⚠️ AskUserQuestion Guard

CRITICAL: After EVERY AskUserQuestion call, check if answers are empty/blank. Known Claude Code bug: outside Plan Mode, AskUserQuestion silently returns empty answers without showing UI.

If answers are empty: DO NOT proceed with assumptions. Instead:

1. Output: "⚠️ Questions didn't display (known Claude Code bug outside Plan Mode)."
2. Present the options as a numbered text list and ask user to reply with their choice number.
3. WAIT for user reply before continuing.

0. Classify Chaos

AskUserQuestion — single call:

"Accidental or deliberate?"

• Accidental (crisis): something broke badly — production down, data corruption, cascading failure → faster gates
• Deliberate (innovation): intentional disruption — novel architecture, paradigm shift, no prior art → exploratory gates

Set mode = accidental|deliberate.

1. Frame the Void

State what is NOT known (not what is known):

• No visible cause-effect relationships
• No best practice applies
• No expert can say "do X"

Identify: what is the minimal safe action to impose ANY constraint on this space?

2. Act-Gate Loop

Repeat until structure emerges or transition triggered:

Act

• Propose ONE minimal action (smallest possible intervention)
• State expected signal: "if this works, we'll see X"
• isolation: worktree for containment if code changes involved

Gate

AskUserQuestion: "Ready to act? Describe outcome after: [action]"

Wait for human response. Gate fires after EVERY action — no batching.

Sense

Interpret the response:

• Signal received → update constraint map
• No signal → action was too small or wrong dimension
• Negative signal → constraint hardened in wrong direction

Log: action N: [what] → [outcome] → [constraint discovered]

3. Structure Check

After each gate, assess:

Constraint type now:
- Still absent → continue act-gate loop
- Enabling (patterns visible) → TRANSITION to /probe
- Governing (experts applicable) → skip /probe → /frame-problem → /investigate
- Rigid (process known) → skip /probe → /frame-problem → execute
- Misclassified → TRANSITION to /frame-problem

4. Exit + Handoff

AskUserQuestion: "Which transition? [Probe — structure emerged / Frame-problem — misclassified / Troubleshoot — crisis was failure]"

Populate handoff template:

• Structure emerged → experiment-to-probe-llm.md
• Re-classify → experiment-to-frame-problem-llm.md
• Crisis / failure → invoke /troubleshoot with action log as context

Observability: ⚡ Experiment → {mode} → {N} acts → {N} gates → {structure} → TRANSITION

Invariants

• NO run_in_background — fully foreground
• Gate after EVERY action — human decides at each step
• isolation: worktree for containment only, never for parallelism
• Exit to /probe when enabling constraints emerge (Chaotic → Complex)

Refs

• reference.md — protocols, gate patterns, transition triggers