Teaching AI How to Think, Not What to Do
This repository contains the foundational papers establishing Framework Injection — a new paradigm for human-AI interaction that transfers complete domain-specific reasoning methodologies from human experts to AI agents.
Prompt Engineering ⊂ Context Engineering ⊂ Framework Injection
(commands) (context mgmt) (reasoning transfer)
"give a recipe" "equip a kitchen" "teach someone to cook"
"Digital Craftsmanship: A Semiotic Framework for Human-AI Interaction Beyond Prompt Engineering"
The philosophical framework — from Plato's evocative language through Aristotle's logic, Medieval grammar, to Digital Craftsmanship. Introduces the 7 operational layers, CAAJ 2.0 cycle, and Linguagem Plena as antidote to "digital pidgin."
- File:
artesanato-digital.tex
"Framework Injection: Teaching AI How to Think, Not What to Do — A Comparative Analysis of Cognitive Transfer Paradigms"
The differentiating concept — formal definition, five-type taxonomy, six theoretical foundations (Peirce, Vygotsky, Sweller, Wittgenstein, Polanyi, Greenberg), comparative analysis against DSPy/LLM-ACTR/Context Engineering, and research agenda.
- File:
paper.tex
| Type | What it transfers | Example |
|---|---|---|
| Declarative | What to know | Legal source hierarchy: Constitution > Statute > Case Law |
| Procedural | How to reason | Clinical differential diagnosis protocol |
| Evaluative | How to judge | Financial due diligence scoring rubric |
| Ethical | What to refuse | LGPD/HIPAA compliance fences |
| Compositional | How to combine | CAAJ 2.0 (Generate → Refute → Improve → Shield) |
| Theory | Author | Connection |
|---|---|---|
| Semiotics | Peirce (1931) | Prompt as sign, LLM as interpretant |
| Linguistic Universals | Greenberg (1963) | Agent/Action/Patient as prompt primitives |
| Zone of Proximal Development | Vygotsky (1978) | FI as scaffolding |
| Cognitive Load Theory | Sweller (1988) | FI reduces extraneous load via schemas |
| Language Games | Wittgenstein (1953) | Each framework defines valid reasoning moves |
| Tacit Knowledge | Polanyi (1966) | FI externalizes expert tacit knowledge |
This work is part of a broader research ecosystem:
- STEER — 16 algebraic operations for embedding-space retrieval (DOI: 10.5281/zenodo.19325724) | Code
- IMI — Cognitive memory for AI agents (DOI: 10.5281/zenodo.19325745) | Code
- IMI Theory — Memory as Field: persistent homology for metacognition (DOI: 10.5281/zenodo.19341740)
- SYMBIONT — 8 biological swarm patterns for multi-agent coordination (DOI: 10.5281/zenodo.19325749) | Code
Renato Aparecido Gomes Independent Researcher, São Paulo, Brazil ORCID: 0009-0005-7380-9876
Apache 2.0 — see LICENSE.
@article{gomes2026frameworkinjection,
title={Framework Injection: Teaching AI How to Think, Not What to Do},
author={Gomes, Renato Aparecido},
year={2026},
note={Preprint}
}