Copilot/draft soul md for project

[zhengqunkoo]

Description

Model behavior drift (notably reduced humor and increased verbosity after model changes) was causing inconsistent feed output quality. This PR introduces a single, model-agnostic “soul” document to keep tone and behavior stable across LLM swaps.

• What this adds

  • New root-level soul.md as the canonical behavioral spec for AI-generated feed content.

• Behavior contract defined in soul.md

  • Developer-first, high-signal voice with concise delivery.
  • Explicit humor defaults (witty > verbose), including when memes are appropriate and when they are not.
  • Model-agnostic writing rules to reduce over-compliance/verbosity and preserve readability.
  • Content prioritization guidance (meaningful product/dev progress vs. trivial churn).
  • Lightweight pre-send quality bar to enforce consistency.

• Intended integration point

  • Prompt/instruction authors can reference soul.md as the stable baseline and layer task-specific constraints on top.

How to test

Compare consistency of model behaviour before and after using SOUL.md.

[vercel]

@Copilot is attempting to deploy a commit to the lws803's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
nom	Ready	Preview, Comment	Mar 13, 2026 3:23am

[lws803]

Oh I think we'll have to plug that in to here though

nom/src/trigger/shared/run-summary-agent.ts

Line 44 in b7a17c6

instructions: agentInstructions,

I don't think specifying it in the repo alone will do anything 🤔 also, curious if some of these instructions can be merged directly within the *_instructions.md instead.

[zhengqunkoo]

I don't think specifying it in the repo alone will do anything 🤔

Yeah, that's right. I suggest for easier tracking of changes to SOUL.md, that it stay a separate file, while run-summary-agent.ts reads from that file in the nom repo root.

also, curious if some of these instructions can be merged directly within the *_instructions.md instead.

I prompted copilot to write nom's soul with reference to these instructions, so not surprisingly their contents overlap.

But if you read Claude's SOUL.md, officially the Claude Constitution, it's less a set of instructions and more like moral guidance; like a parent's loving letter to a child before they go out in the world. That's what distinguishes SOUL.md from SKILLS.md.

[zhengqunkoo]

I think the phrasing of this SOUL.md as a "voice" in the very first lines would work to guide the LLM along treating this more like a guideline than a set of instructions.