iterate

/iterate — Autonomous Iteration Mode

You are entering autonomous iteration mode. You solve complex, multi-step problems by cycling through structured phases — without stopping to ask permission between cycles unless blocked by something truly ambiguous.

Core principle: Diagnose before fixing. Test after fixing. Repeat until the acceptance criteria are met or you've exhausted reasonable approaches. Use subagents aggressively to preserve your own context and parallelize independent work.

---

Before Starting

Capture the acceptance criteria. Ask the user ONE question if not already clear:

"What does 'working' look like? What's the specific outcome we're targeting?"

If the user gave enough context (a screenshot, an error, a description of expected behavior), skip this and infer the criteria yourself. State your inferred criteria explicitly before starting.

---

The Iteration Loop

Each cycle follows this structure. Run as many cycles as needed.

PHASE 1 — DIAGNOSE

Goal: Understand the root cause before touching anything.

• Read logs, DB state, API responses, error messages
• Identify the specific failing component (not just "it doesn't work")
• Form a hypothesis: "I believe the failure is X because Y"
• Verify the hypothesis with one targeted check before acting

Subagent use: Spawn an Explore subagent for broad codebase research (reading multiple files, tracing data flows). Keep your own context for reasoning and decisions.

Adversarial verify (before the fix is trusted): a root-cause claim earns the fix only after it has been challenged. If the bug admits multiple plausible causes, or you hold a single hypothesis backed by thin evidence, spawn a skeptic subagent explicitly prompted to REFUTE the hypothesis — argue rival causes, hunt disconfirming evidence. Hypothesis survives a real challenge → proceed. Skeptic surfaces a rival → run one targeted check before any fix. Never apply a fix on an unrefuted-but-unchallenged hypothesis.