Prompting & Discipline
The April 2026 research converged on one finding: the harness — CLAUDE.md, hooks, skills — matters more than the prompts. And the CLAUDE.md itself can be hardened using persuasion-psychology patterns that are empirically validated on LLMs.
What you’ll find here
- Hightower’s Cialdini playbook — authority / commitment / social proof / inverted scarcity applied to CLAUDE.md, with rationalisation tables and 13 red flags
- ZIRU’s 90%-missing framing — why most users are stuck prompt-engineering instead of harness-engineering
- The Agent Harness vocabulary — Njenga’s community-standard term for CLAUDE.md + hooks + skills + MCPs
- Karpathy’s CLAUDE.md principles — what each rule actually fixes
- The 235-skills audit — Rezvani’s field report on skill-library rot and the 4 categories that compound
Featured in this category
- Superpowers: Cialdini’s Psychology Hack for LLMs — $28K Wharton study, compliance 33% → 100%
- Anthropic runs hundreds of skills — I published 235, 12 run weekly — the audit methodology with 4 complete SKILL.md templates
- Agent Harness — what pros understand — the community framing
Why this category matters more on Opus 4.7
Opus 4.7’s literal-interpretation behaviour amplifies both the authority and commitment effects documented in the Wharton study. Weak CLAUDE.md files produce visibly worse output on 4.7 than on 4.6; well-structured ones produce noticeably better output. The wording change from “should” to “MUST” is a free upgrade on every prompt your team runs.