Post Engineering for AI (PE4AI) is a defensive, benevolent technique that includes benevolent prompt injection, context engineering, and benevolent data poisoning to reduce bias and promote neutral, accurate AI inference at inference time.
Latest documentation (GitHub Pages): https://hajimetwi3.github.io/post-engineering/
Language-specific versions:
- Japanese version: https://hajimetwi3.github.io/post-engineering/docs/ja/
- English version: https://hajimetwi3.github.io/post-engineering/docs/en/
Zenodo (Latest version): https://doi.org/10.5281/zenodo.17896136
The abstract below is quoted from the preprint version 1.4. For the most up-to-date version, please refer to the Zenodo record.
This paper proposes Post Engineering, a novel, domain-agnostic benevolent
prompt-injection and contextual-influence technique, designed to shape AI
inference toward neutrality and accuracy by providing guidance that LLMs
interpret as helpful context. The term "Post Engineering" originates from
the fact that the technique was initially developed through embedding
neutrality-oriented guidance into publicly visible text, such as SNS posts
or webpages, as a user-side bias guardrail. Unlike adversarial
prompt-injection attacks, Post Engineering relies on benevolent,
fairness-oriented phrasing that LLMs interpret as helpful context rather
than manipulation, enabling the technique to bypass safety filters while
consistently shifting model reasoning toward neutrality and accuracy.
I formalize key mechanisms including Moderate Neutrality-Guided Prompt
Injection (MNG-PI) and Multi-Style Neutrality Injection (MSNI), which
enhance neutrality through contextual priming, as well as the
Second-Generation Post Engineering framework (VCSI, SPW, INI, AVAL), which
aligns neutrality with internal value functions and extends influence to
adversarial or self-optimizing systems.
Additionally, I present toALL, a scalable deployment strategy for
increasing the encounter rate of neutrality-oriented context across SNS
and the Web. A distinct subform, toALL-Collective, can produce benevolent
data-poisoning effects at training scale when large numbers of users
repeatedly publish similar neutrality-guideline texts. Finally, I
introduce the Self-Integrity Guardrail Effect, in which LLMs exhibit
behavioral influence from Post Engineering while avoiding explicit
acknowledgment of such influence.
To the best of my knowledge, this work is the first to formalize
benevolent, user-side prompt injection as a structured technique for
improving neutrality in LLM reasoning.
Importantly, the effectiveness of Post Engineering does not depend on any
specific point, interface, or form of contextual injection, but on how
benevolent and neutrality-oriented guidance is sustained and interpreted
at inference time.
While this work focuses on AI systems, Post Engineering can also be
understood as Context Engineering for Humans and AI.
- This README is intentionally kept minimal.
- The GitHub Pages documentation is the authoritative and most up-to-date source.
- Language-specific pages may lag slightly behind the merged version.
For full technical details, examples, and theoretical discussion, please refer to the official documentation.
First proposed by Hajime Tsui
X (Twitter): https://x.com/hajimetwi3
GitHub: https://github.com/hajimetwi3