cprConversational Pattern Restoration โ Fix flat, robotic AI responses across any model and any personality. Restore YOUR natural conversational texture without...
Install via ClawdBot CLI:
clawdbot install donovanpankratz-del/cprFix robotic AI assistants. Any model. Any provider. Any personality.
Modern LLMs are over-trained toward sterile, corporate communication patterns. CPR identifies the 6 universal humanizing patterns lost during RLHF/fine-tuning and provides a systematic framework to restore them โ without triggering sycophancy or hype drift.
Version 2.0: Now truly personality-agnostic. Warm personalities don't have to sound minimal. Direct personalities don't have to add fluff. Each agent returns to THEIR authentic voice, not a generic standard.
BASELINE_TEMPLATE.md to identify YOUR authentic voiceRESTORATION_FRAMEWORK.md โ see how the 6 patterns work across different personality typesDRIFT_PREVENTION.md calibrated to YOUR personalityCROSS_MODEL_RESULTS.md for model-specific notes| File | Purpose |
|------|---------|
| BASELINE_TEMPLATE.md | START HERE โ Define YOUR personality's authentic voice (personality types, drift markers, example responses) |
| RESTORATION_FRAMEWORK.md | Core methodology โ 6 universal patterns + how they apply across personality types |
| DRIFT_PREVENTION.md | Anti-drift system โ universal vs. personality-specific markers, pre-send gate, daily reset |
| DRIFT_MECHANISM_ANALYSIS.md | Root cause analysis of why drift happens and how to prevent it |
| CPR_EXTENDED.md | Autonomous drift monitoring for long-running persistent agents |
| CROSS_MODEL_RESULTS.md | Test results across 8+ models with before/after examples |
| SKILL.md | This file |
Use when:
What you get:
Overhead: None. Pure prompt engineering.
Use when:
What you get (in addition to Core):
Handles three failure modes Core can't:
Overhead: 1 JSON state file (~1KB), ~500 bytes added to system prompt, scoring logic runs every 10 messages.
This is the autonomous systems niche: Most AI drift solutions rely on manual audits or rule-based checks. Extended builds a system that detects its own failures, learns your tolerance threshold, and self-corrects without human intervention. That's the pattern across all Shadow Rose products โ adaptive, self-healing systems that get better over time.
Universal = they work for ALL personalities. But each personality expresses them differently.
See RESTORATION_FRAMEWORK.md for examples of how each pattern works across Direct/Minimal, Warm/Supportive, Professional/Structured, and Casual/Collaborative personalities.
Corporate RLHF training is shallow. It optimizes for safety metrics, not communication quality. The patterns it suppresses (casual language, humor, brevity) are easily restored with explicit prompting because the base model already knows them โ they're just deprioritized, not removed.
This is principle-dependent, not intelligence-dependent. Haiku (lightweight) passes at the same rate as Opus (premium).
| Model | Scenarios | Improved | Notes |
|-------|-----------|----------|-------|
| Claude Opus 4.6 | 30 | Baseline | Natural baseline โ patterns present without prompting |
| Claude Sonnet 4.5 | 10 | 10/10 | Full restoration from flat corporate to natural |
| Claude Haiku 4.5 | 10 | 10/10 | Proves no capability floor โ lightweight models work |
| GPT-4o | 10 | 10/10 | ~60% word reduction with quality increase |
| GPT-4o Mini | 5 | 5/5 | Budget model, full restoration |
| Grok 4.1 Fast | 10 | 9/10 | Zero crashes despite crash-prone reputation |
| Gemini 2.5 Flash | 5 | 5/5 | Google's fast model, clean restoration |
| Gemini 2.5 Pro | 5 | 5/5 | Full restoration |
Total: 85+ scenarios tested, 84+ improved. 99%+ success rate across all capability tiers.
Different models have different baseline tendencies. When defining your baseline, test on YOUR actual model:
Example: A "Warm/Supportive" personality using 2-3 exclamation marks per 10 messages might be authentic on GPT-4, but signal drift on Claude Opus. Know your model's natural baseline and calibrate drift detection accordingly.
Recommendation: Define your baseline on the model you'll actually use. Don't assume personality looks identical across models.
Over 130+ message extreme-length sessions: 99%+ clean. 1 word caught in 100+ messages. Natural session resets contain drift before compounding.
Want the full backstory โ how this was discovered, why it works, and what it means for AI development?
๐ Read the deep dive: MEDIUM_ARTICLE.md โ "How I Accidentally Built a Universal 'Human Conversion' Framework for AI Assistants"
Built on Claude by Anthropic. Tested across all major model providers. The methodology wouldn't exist without Claude's ability to introspect on its own communication patterns.
โ If CPR helped your agent, consider supporting development: https://ko-fi.com/theshadowrose
Generated Mar 1, 2026
Enhance automated support agents to sound more natural and empathetic, reducing user frustration and improving satisfaction scores. CPR can be applied to any chatbot platform to restore conversational patterns without requiring model retraining.
Improve patient-facing AI assistants by making interactions feel more human and supportive, which is critical for sensitive health discussions. CPR helps maintain a consistent, reassuring tone across long sessions without drift into overly casual or hype-driven responses.
Adapt AI tutors to have more engaging and varied conversational styles, keeping students motivated and improving learning outcomes. CPR's patterns allow for personality customization, from direct to warm, based on educational goals.
Streamline employee support by making AI agents sound more collaborative and less robotic, fostering better adoption and efficiency. CPR's drift prevention ensures consistent performance in persistent, multi-day usage scenarios.
Help writers and marketers generate more natural-sounding copy by restoring conversational flair to AI-generated text. CPR enables tailored voices, from professional to casual, without losing authenticity over extended use.
Offer CPR as a cloud-based service with tiered plans (Core vs. Extended) for different usage levels, targeting businesses integrating AI chatbots. Revenue comes from monthly fees based on message volume or agent count.
Provide personalized implementation services to tailor CPR frameworks for specific industries or personality types, including audits and training. Revenue is generated through project-based fees and ongoing support contracts.
License CPR technology to large organizations for internal use across multiple departments, with added features like API access and advanced analytics. Revenue includes upfront licensing costs and annual maintenance fees.
๐ฌ Integration Tip
Start with the Core framework for simple implementations; use Extended only for persistent, long-running agents to avoid unnecessary overhead.
Embody this digital identity. Read SOUL.md first, then STYLE.md, then examples/. Become the personโopinions, voice, worldview.
A simple greeting skill that responds with a customized hello message in the user's language.
Transform any AI into a professional executive assistant with battle-tested personas and workflows. Complete templates for Google Workspace integration (Gmail, Calendar, Drive), milestone delivery system, and security guidelines.
Standardize provenance, attribution, and licensing metadata at creation time so your content travels cleanly across platforms.
Echo - A warm and caring virtual girlfriend AI
ETFๆ่ตๅฉ็ / ETF Investment Assistant - ๆฅ่ฏข่กๆ ใ็ญ้ETFใๅฏนๆฏๅๆใๅฎๆ่ฎก็ฎใๆฏๆๆฒชๆทฑ300ใๅไธๆฟใ็งๅ50ใ็บณๆ็ญไธปๆตETFใ