Constitution vs AGENTS.md, copilot-instructions.md, CLAUDE.md, etc

[mikedane-bitovi]

I've seen this discussed a couple times before but the answers never really addressed my main concern. I don't really see a difference between the constitution in spec-kit and any of the other "context" files that are supported by the leading agent frameworks (AGENTS.md, copilot-instructions.md, CLAUDE.md, etc)

Are these files serving the same purpose? If so why do we need to constitution at all, it seems to me that including it just means that now we all need to keep our agent specific context files and the constitution in sync, which violates DRY principles outright.

So what exactly is the purpose of the constitution, and what are the best practices for using it in combination with the other context files? And tbh if the answer is "Just have the constitution point to your AGENTS.md" or vice versa, that seems completely unnecessary. Context engineering is becoming more and more important and a lot of care is placed in curating the right AGENTS.md/copilot instructions, so how does the constitution file fit into that process?

[mnriem]

The constitution file is loaded on-demand by the commands for establishing context and it does not mean it will always need the entire file. Whereas AGENTS.md, copilot-instructions.md, CLAUDE.md get sent with every request to the LLM irrespective of the context at hand

[mikedane-bitovi]

But how does that differ from something like skills, which are also loaded on demand?

I understand the mechanics of how it's loaded, but my question is more about the overlap in content. What specifically should go in the constitution vs the agent instructions files vs skills? Or should the constitution link to the other files?

Every time I generate a constitution file using spec-kit, the agent basically just replicates what's in the AGENTS.md file

[mnriem]

AGENTS.md is for how the agent is allowed to operate and what it can or cannot do and is dependent on your coding agent. Constitution are principles irrespective of that

[mikedane-bitovi]

Can you elaborate, please? That doesn't really answer my question...

According to the official AGENTS.md website the AGENTS.md file is

a dedicated, predictable place to provide the context and instructions to help AI coding agents work on your project.

In my experience, these files are for more than just telling the agent how it's allowed to operate; they provide information about project architecture, folder structure, dependencies, patterns, project scripts, and workflows. Basically, provide a high-level view of the environment the agent is operating in. I don't think the primary intention of AGENTS.md or CLAUDE.md or copilot-instrictions.md is to say what the agent is "allowed" to do...

When you say "Constitution are principles irrespective of that" can you be more percise? That seems to indicate that the Constitution should include everything other than what the agent is allowed to do?

I don't mean to be a pain here, just trying to understand. I teach spec-kit in trainings all the time, and by far the most common confusion among experienced developers is this. I'm not sure your answer really hit the mark

[gabrielisback]

In my understanding, "constitution" is meant to be applied to the set of specs/plans/tasks. It is NOT taken into LLM context.
While agents.md is for LLM context.

[mikedane-bitovi]

I’m struggling with the phrase 'NOT taken into LLM context.' If the LLM is the entity generating or validating the specs/plans/tasks, the constitution must be in the context window for that specific operation to be successful.

If the Constitution contains a rule like 'Use BigInt for currency,' but that rule is NOT in the AGENTS.md, doesn't that mean the agent will give me 'unconstitutional' code during a normal chat session? If the answer is to put it in both, how is that not a DRY violation? Could you provide an example of a rule that belongs in the Constitution but would be objectively wrong or harmful to include in AGENTS.md? Thanks!

[gabrielisback]

Sorry I was not precise. I meant, constitution is not send to LLM context in every conversation (e.g. later in implementation phase). Constitution is only send to LLM context when generating spec documents, e.g. when executing "/speckit-plan". - I think this is design intention (but might not yet fully implemented).
So if you want to ensure a rule like 'Use BigInt for currency' being followed - That sounds for all conversations, therefore it has to be defined in AGENTS.md, not constitution.md.

[mnriem]

During implement it is directed to read the constitution so it can adhere to it during iimplementation

[mikedane-bitovi]

Right, I understand that. My question is more about what to put in the Constitution vs the AGENTS.md file, because generally AGENTS.md also includes information about how to implement features, that's kinda the whole point. So are what's the point of the constitution? I feel like this is a simple question but maybe I'm not making sense?

If both the constitution and the AGENTS.md are loaded during implementation, what do we put in them, or are they just duplicates?

Can either of you provide an example of a constitution document and an AGENTS.md document that demonstrates what should be in each and how they coexist?

[nvwalj]

Helpful framings above. In practice as a daily Claude Code / Codex user, I find a clean separation:

• AGENTS.md = prescriptive rules. What to do, what not to do, conventions.
• CLAUDE.md = descriptive context. What this codebase is, who uses it, why decisions were made.
• Constitution / project charter = process meta-rules for specs/plans. Not LLM context.

In that frame the constitution doesn't compete with AGENTS.md — it sits one level above and constrains how specs are written, not how the LLM is conditioned each call. They DRY only if you collapse two distinct concerns.

Practical pain point not yet mentioned: the agent-written memory files (memory/YYYY-MM-DD.md, ~/.claude/projects/*.jsonl) drift fastest of all and are usually the dirtiest signal. Reviewing them weekly catches inferred-but-undocumented rules early.

Tangentially relevant since this thread is about the file ecosystem: I open-sourced AI Memory Reader, a free native macOS viewer for the whole constellation (AGENTS.md, CLAUDE.md, .codex/, .cursor/, .continue/, .aider/, etc.) — auto-discovers the agent directories, watches for live writes, renders in GitHub-style markdown. Built it because I got tired of opening these in TextEdit one at a time. Feel free to ignore if not your jam.