A meta-prompt system for agentic CLIs (Claude Code, Gemini CLI, Codex, OpenCode, Cursor, Aider). It classifies a user's rough task idea into one of nine domains, applies domain-specific best practices anchored in established frameworks (Diátaxis, Bloom's taxonomy, PRISMA, pyramid principle, backward design, schema-first data practices, and others), asks a small batch of clarifying questions, and generates a single structured prompt that a fresh agent session can execute to build a planning scaffold for the work.
Every type module was refined through a controlled multi-agent experiment: 9 end-to-end units, 3 conditions per unit (baseline, simple ablation, targeted augmentation), 27 producer scaffolds, 9 blind evaluators on a 5-criterion rubric. Eight proposed augmentations produced directional wins and were merged; one was rejected because the experiment showed it actively hurt scaffold quality. The full audit trail is reproduced in the appendices of the attached report.
Included in this release:
- Nine A/B-tested type modules: includes three identical harness orchestrator files, the autonomy and compaction-resilience guardrails, the beginner-oriented README, a contribution guide, and a PR template with mandatory A/B evidence, plus an agent-guided workflow for adding new types.
- promptprimer-report.pdf: 16-page creation and refinement report with full experimental methodology, per-unit results, blind-evaluator summaries, and applied changelog.
MIT licensed. knowledge/bestpractices.md is separately sourced from Anthropic's documentation and governed by Anthropic's terms; see the attribution section of the README.