Spike

Structured technical investigation to reduce uncertainty. Answer specific questions, not "explore X."

Key Principles

• Every spike answers a specific, measurable question
• Strict time-box (default 2h) — produce a decision at the end, even if incomplete
• Output is a decision (GO/NO-GO/PIVOT/MORE-INFO), not a general exploration
• Create follow-up tasks from findings, not just a report

When NOT to Use

• Multi-day evaluations — Use deep-research for comprehensive technology evaluations with stakeholder reports
• Already know what to build — Skip straight to implementation; spikes are for reducing uncertainty, not planning known work
• Bug investigation — Use error-handling or /fix; bugs have reproduction steps, spikes have open questions

Quick Start Checklist

1. Define the question clearly (what are we trying to learn?)
2. Set a time-box (hours, not days — use deep-research for multi-day)
3. Identify evaluation criteria upfront
4. Investigate with focused experiments or prototypes
5. Checkpoint at 50% — assess progress, decide if pivoting
6. Produce mandatory conclusion: GO / NO-GO / PIVOT / MORE-INFO

Mandatory Outputs

• Decision: GO, NO-GO, PIVOT, or MORE-INFO
• Evidence: What was tested, results observed
• Follow-up tasks: Created in ohno if GO
• Report: Saved to .claude/spikes/[date]-[slug].md

References

Runtime Notes

• Claude Code: dispatch yokay-spike-runner via the Task tool with subagent_type: "pokayokay:yokay-spike-runner".
• Codex: there is no subagent dispatch. Execute the agent's role inline — read the corresponding agents/yokay-<name>.md and follow its Behavioral Defaults, Critical Rules, and Output Contract directly in the current session.