Strategy Synthesis Skill

Steps

1. Check target files exist:

   • Check KAGGLE_DIRECTION.md / claudeSummary.md / daily_reports/ / submit/SUBMISSIONS.md / workspace/
   • If any missing, mark "Not found" and pass through to agent

2. Invoke competition-strategist agent:

   • subagent_type: competition-strategist
   • In the prompt, specify:
     • Competition name and deadline (from KAGGLE_DIRECTION.md)
     • Phase detection must be the first thing (time-based + milestone-based; warn on misalignment)
     • All subsequent proposals are restricted to that phase's "do" list (see KAGGLE_DIRECTION.md "Phase Guidance")
     • $ARGUMENTS as additional focus if given (e.g., "focus on ensemble strategy", "emphasize LB shake risk")
     • Output target: append to the "Strategy & Roadmap" section of today's daily report daily_reports/YYYYMMDD.md

3. Append agent output to today's daily report:

   • Append to the "Strategy & Roadmap" section of today's daily report daily_reports/YYYYMMDD.md (never create a separate strategy_*.md file — records are consolidated in the daily report)
   • If today's daily report doesn't exist yet, create it via the /daily-report steps first, then append
   • Summarize key proposals (phase detection, safe+bold, unexplored, risks) for the user
   • Surface misalignment warnings first (e.g., "Entered mid phase but no working baseline")

4. Propose next actions:

   • Pick 1-2 of the proposed next moves to recommend for tomorrow
   • Suggest /exp-new for new experiment folders or /survey-papers for further research as needed

When to Use

• Weekly: as a regular checkpoint
• CV plateaus: to break out of local optima
• After LB shake: to step back and identify the cause
• New data drop: to re-evaluate strategy
• 1 week before deadline: to optimize remaining resource allocation

Notes

• Agent output is fact-based. Inferences are marked "Inference:"
• "Best X" claims must always cite source (experiment folder, daily report date)
• Don't blindly execute agent suggestions — treat them as input for user judgment