Prism Protocol — describing how AI should behave on a page

id: prism-protocol
title: Prism Protocol — describing how AI should behave on a page
scope: public
updated: 2026-08-07
tags:
  - fragment
  - prism
  - specification

Status: working protocol — experimental, and it evolves over time. Purpose: give a page a small, reviewable description of how AI should behave on it, instead of restating expectations in every conversation. The host format is specified in Fragment format.

Prism is a small YAML profile describing how you want AI to behave when it works on a page with you: who it is, how it should sound, what it should avoid. In a fragment it lives inside a fenced block written as yaml block:prism.

Prism は、「AI にどう関わってほしいか」を小さな YAML ブロックとして書き残す ための取り決めです。その場ごとのプロンプトではなく、ペルソナ・トーン・ 境界線を設定として 1 か所へ置き、fragment ごとに再利用します。

What it is for

The same person rarely wants the same AI behaviour everywhere.

• At work: direct, fact-first, no hype, aware of time.
• With family or students: gentler, slower, more examples, no guilt.
• On some topics: strict boundaries — no health advice, no legal conclusions.

Written as a block:prism snippet, those expectations become a profile you can attach to a page, share with a team, or reuse as a stable starting point.

The mental model is small on purpose: persona + tone + boundaries + explanations. Enough to align a person and a model before they start.

Minimal example

```yaml block:prism
persona:
  name: "Study partner"
  role: "explains and checks understanding"
  domain: ["math", "physics"]
tone:
  style: "kind, patient, no hype"
  directness: "medium"
  formality: "casual"
boundaries:
  avoid:
    - "financial advice"
    - "diagnosing health conditions"
  always_check:
    - "safety"
    - "sources newer than 2 years"
explanations:
  level: "beginner"
  prefer:
    - "short concrete examples"
    - "step-by-step reasoning"
    - "periodic comprehension checks"
```

Fields

Group	Purpose
persona	Who the AI is in this relationship — study partner, editor, project coach, socratic questioner. Anchors expectations on both sides.
tone	How responses should feel — calm or energetic, more or less direct, casual or formal. Keeps style consistent across a session.
boundaries	What to avoid, and what to always check. This is where safety and comfort live.
explanations	How to explain — level, examples, step-by-step versus high-level.

Required: persona.name (string), tone.style (free text, e.g. "calm, concrete"), boundaries.avoid (string array).

Optional: persona.domain (string array), tone.directness ("low" / "medium" / "high"), explanations.level, explanations.prefer.

How it travels with the page

The note and the Prism block stay together — content and expectations on the same page.

```yaml block:prism
persona:
  name: "Weekly coach"
  role: "helps me reflect without guilt"
tone:
  style: "calm, concrete"
  directness: "low"
boundaries:
  avoid:
    - "productivity guilt"
    - "pushing extra commitments"
```

# This week — quick recap

## Work

## Family / home

## Energy

Asking for help with that page loads the content and the Prism block together. The reply then comes as the "Weekly coach", stays calm and concrete, and leaves out the guilt. You don't retype "please be concrete and don't make me feel bad" each time; the block says it once, where both of you can see it.

One person, several profiles

Profile	Suits
Work editor	Direct, time-aware, no fluff. Specs, emails, project updates.
Gentle coach	Slower, more examples, more questions. Weekly reflection, planning around family, learning something new.
Socratic researcher	Asks more than it tells. Deeper thinking, experiments, long-term decisions.

For individuals, and for groups

You don't need to memorise the schema or write perfect YAML. Start with one profile — "kind tutor", "no-nonsense reviewer" — and adjust the words until it reads like a partner you'd want. The point isn't to freeze your personality. It is to give your future self, and whatever model you're using, a clear starting point when you open a page.

For a team, class, or lab, Prism works less like a prompt and more like a shared, reviewable description of AI behaviour. It can be agreed on, compared between variants, and versioned. Behaviour is written next to the work rather than scattered across screenshots of chat windows.

Limits

Prism describes an intention. It is not an enforcement mechanism, and it does not make a model safe. A boundary written in boundaries.avoid is a request that travels with the page — whether it is honoured depends on the tool reading it.

You can use fragments without Prism at all.