OPINIONS.md Pattern

What it is

A compact, living map of what a person actually believes, maintained by an agent through a daily cron job. Unlike an AGENTS.md or a knowledge base, OPINIONS.md is not a cookbook of implementation details. It is a structured file of durable opinions, principles, taste, values, critiques, predictions, and tradeoffs — extracted from a person's public writing and activity.

The pattern was first described by Kun Chen in May 2026, who uses Hermes Agent (or OpenClaw) with GPT 5.5 to maintain his own file.

Why it matters

Most agents fail because they do not understand what their user cares about. A generic system prompt cannot encode the subtle, evolving texture of a person's judgment. OPINIONS.md solves this by:

1. Making taste legible. It converts scattered posts into a structured belief map that both humans and agents can read.
2. Enabling alignment. Agents that read OPINIONS.md before acting can produce outputs that match the user's actual views rather than generic consensus.
3. Supporting self-reflection. The file includes a watchdog that flags opinion drift, contradictions, and factual risks — turning the agent into a mirror rather than just a tool.

Key points

• Filter is more important than extraction. The critical part of the cron prompt is not "read my posts" but what to leave out: jokes, dunks, throwaway replies, one-off reactions, ambiguous posts without context, and technical walkthroughs that do not reveal broader opinions.
• The file is allowed to reorganize itself. Section structure should not be static. After each update, the agent asks whether the current structure is still the clearest way to organize the opinions.
• Watchdog pass runs even when no new sources exist. It checks for opinion drift (new posts that contradict or refine existing opinions) and factual risk (claims that may have become stale as the world changes).
• Self-criticism is a feature. Kun Chen reports that asking the agent "which of these opinions look objectively wrong" produced extremely useful feedback — pointing out overgeneralizations, imperfect analogies, and claims that needed qualification.

Evidence across sources

Design principles

1. Durable over ephemeral. Extract opinions stable enough to matter, not momentary reactions.
2. Evidence-linked. Each opinion cites the source post or essay that supports it.
3. Self-reorganizing. The structure should improve over time, not just append.
4. Uncertainty-preserving. Possible views that do not establish durable opinions are mentioned as weak signals at most.
5. Watchdog-guarded. Drift detection and factual-risk scanning run on every pass.

Open questions

• How does OPINIONS.md differ from a well-maintained AGENTS.md or CLAUDE.md in practice?
• What happens when a person's opinions are genuinely contradictory across contexts?
• Can this pattern scale to team or organizational belief maps?

Prompts for witness

• If an agent maintained your OPINIONS.md, what categories would dominate the file?
• Which of your publicly expressed opinions would you be most surprised to see summarized back to you?
• What is the difference between "taste" and "bias" in this framework?

Related

• What Makes a Good AGENTS.md
• Pirate and Architect — 2026 年软件工程的两种角色
• Fat Skills, Fat Code, Thin Harness
• AI 创始人的品味指南