Response Pages Style Guide

Response pages describe interventions, policies, and technical approaches that address AI risks. They explain how something works and assess its effectiveness.

Prerequisite: All response pages must follow the Common Writing PrinciplesE726Shared writing principles referenced by all domain-specific style guides. Three pillars: epistemic honesty (hedge uncertain claims, use ranges, source confidence levels), language neutrality (avoid...0 — epistemic honesty, language neutrality, and analytical tone. The objectivity rating dimension measures this.

Page Type Detection

Response pages are at: /knowledge-base/responses/**/*.mdx

Required Frontmatter

---
title: "Response Name"
description: "One sentence explaining what this response does and its key mechanism."
quality: 60  # 0-100
readerImportance: 70  # 0-100
lastEdited: "2026-01-28"
---

Required Sections

1. Overview (2-3 paragraphs)

What is this response and why does it matter?

2. Quick Assessment Table

## Quick Assessment

| Dimension | Assessment | Evidence |
|-----------|------------|----------|
| Tractability | Medium | Requires significant research investment |
| Scalability | High | Applies to most foundation models |
| Current Maturity | Low | Early research stage |
| Time Horizon | 5-10 years | Needs fundamental advances |
| Key Proponents | Anthropic, DeepMind | Active research programs |

3. How It Works

Technical explanation with diagram:

## How It Works

<Mermaid chart={`
flowchart LR
    A[Input] --> B[Safety Layer]
    B --> C{Check}
    C -->|Pass| D[Output]
    C -->|Fail| E[Block/Modify]
`} />

[Detailed explanation of mechanism]

4. Risks Addressed

## Risks Addressed

| Risk | Relevance | How It Helps |
|------|-----------|--------------|
| [Deceptive Alignment](/...) | High | Detects hidden goals |
| [Reward Hacking](/...) | Medium | Identifies misspecified rewards |

5. Limitations

What this approach cannot do or gets wrong.

6. Current State

Who is working on this, what progress has been made.

7. Open Questions

Unsolved problems and research directions.

Claude Code Workflows

Creating a New Response Page

Task({
  subagent_type: 'general-purpose',
  prompt: `Create a response page for [RESPONSE_NAME].

  FIRST: Read /internal/response-style-guide/ (Response Style Guide).

  THEN: Research using WebSearch:
  - Academic papers on the technique
  - Lab blog posts and announcements
  - Current implementations and results

  Create at: src/content/docs/knowledge-base/responses/[category]/[name].mdx

  Include:
  1. Overview (2-3 paragraphs)
  2. Quick Assessment table
  3. How It Works (with diagram)
  4. Risks Addressed table
  5. Limitations
  6. Current State
  7. Open Questions`
})

Improving a Response Page

Task({
  subagent_type: 'general-purpose',
  prompt: `Improve response page at [PATH].

  Read /internal/response-style-guide/ (Response Style Guide) first.

  Add:
  1. Quick Assessment table (if missing)
  2. Mermaid diagram
  3. Risks Addressed cross-links
  4. Citations from recent papers

  Use WebSearch for current research.`
})

Quality Criteria

Pages are scored on seven dimensions (0-10 scale). Scoring is harsh - a 7 is exceptional, most content should score 3-5.

Derived quality (0-100) combines subscores with word count and citation bonuses. See CLAUDE.md for formula.

Example

See Mechanistic InterpretabilityApproachMechanistic InterpretabilityMechanistic interpretability aims to reverse-engineer neural networks to understand internal computations, with \$100M+ annual investment across major labs. Anthropic extracted 30M+ features from C...Quality: 59/100 for reference.