Research Report

Use this skill for experiment-result writeups that should sound like the local research-report Claude command.

Workflow

1. Read references/original-claude-command.md when you need the exact framing, CTA, or reply block.
2. Extract the core facts first:
   • total experiment count
   • model and hardware
   • what was automated
   • baseline metric
   • best and worst outcomes
   • any reversals between cheap and expensive lanes
   • whether any runs are still in progress
3. Lead with automation:
   • automated ai research ran these
   • cheap small experiments were intentional so losers get eliminated before scale-up
4. Keep the voice tight:
   • all lowercase sentences
   • direct first person
   • no filler
   • no "i found that" or "we observed"
5. Default structure:
   • opener with n experiments + model + hardware
   • 1-2 lines on the automated system and why the small lane exists
   • markdown table sorted best to worst by vs-baseline result
   • 3-5 bullets on winners, reversals, and failures
   • 1-2 lines on what gets scaled next
6. If the user asks for a blog post instead of an X post:
   • keep the same opener, automation framing, and results table
   • expand the process and reversal sections
   • do not include the CTA unless the user explicitly wants the social version

Output Rules

• bold the best and worst rows in the main table
• use ***, **, *, or n.s. in the signal column
• if a result only survived a cheap screen and was not scaled, say that explicitly instead of pretending it is confirmed
• if timing mattered, include predicted vs actual runtime as part of the story
• if the hook is a reversal between cheap and expensive lanes, lead with that