The repo is a 334k-word corpus of ~105 production system prompts across ~30 AI
products, with no synthesis. This adds the synthesis.
- Part I: twelve rules that multiple independent vendors converged on
(stop conditions, anti-hallucination pairing, parallel-by-default, line
budgets, named filler strings, dependency verification, green-before-done,
root-cause fixes, ask-vs-act bias, search-tool routing, anchored edits,
externalized plans) - each cited to file:line in the corpus.
- Part II: high-value techniques unique to a single product (Manus event
loop, Kiro EARS specs, Claude for Chrome trust boundary, Windsurf
suggestion self-test, VS Code mode selection).
- Part III: prompt-lineage analysis quantifying verbatim cross-vendor reuse
(line-set overlap over 73 files), with the caveat it implies for reading
the convergence evidence in Part I.
- Part IV: a skeleton template implied by the convergent rules.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016F6MR5mdWYJmZJsZzE7rKu
