Markdown in /src

Cluttering the repo with out-dated, very wordy and quickly aging prompts will just confuse any agent tasked with looking at the repo in the future. Keeping context windows down is a real limitation to good LLM output, and this workflow may work completely against it.

- A plan.md describing the project, main abstraction idea, end costumer, and so on is great; but it should be kept minimal and up-to-date with the repo.

- Block comments on top of source-files and functions are great, and already very useful to coding agents. I don't see a value to anything more than what is already typical best practice.

We did put non-code into /src for a very long time: It was heredocs, multiline docs, etc. Actually my preference is to put texts close to code and only fallback to /docs/something.md at a conceptual level. Which is probably what the author proposes, given that he sees markdown as primary interface to code.

One example is SpiderMonkey, which uses these beautiful, long expository comments explaining not only what, but why design choices have been made. https://searchfox.org/firefox-main/source/js/public/RootingA...

If the goal is "locality", you can't get much closer than as a comment.

As far as markdown becoming "source code for agents" under the agentic paradigm, per-directory `AGENTS.md` seems more consistent, at least visually. If agent managed markdown is going to be high-churn, I'd rather it be confined to a single file. Constraints, especially for agents, are good.

I would much rather have significantly more detailed comments inline with code to describe the intention at any given step than just a high level markdown file. We already have PDRs and markdown seems like it would just be an extension of that.

I assume src/ is treated as the source of truth. If we start placing Markdown documentation inside src/, outdated documentation can live alongside the actual source of truth, blurring the boundary between what the system actually does and what the documentation says it does.

This is being done with an agent-first approach in mind. Agents cost a lot to run. Moving markdown to /src will only help AI subscription sellers.

If you are a fan of local LLM, then you also know code is not cheap. It requires substantial amount of investment in infra acquisition and maintenance.

Markdown should be where they always were, in /docs. It does not make a difference for an agent. But it allows to keep /src clean and adhere to zero-token architecture.

In-between all those ephemeral agent sessions, spec, plan and design files, there is something that survives. The src/*.md files.

With Varar you can link small parts (cells) of those Markdown files to your code, so they stay in sync. So you (and agents) can find out what your system does without reading 300 million lines of code.

Full disclaimer: I am the guy who wrote Cucumber - that most people here hate. I eventually lost my love for that tool myself, but I did miss the ability to keep docs and code in sync. Hence Varar - a tool like that with fewer (different?) warts, for the agentic area.

But I don't really like the "LLMs as compiler" metaphor. If you followed that logic to the end, you'd have to "rebuild" your entire project from the spec every time the spec changes. Not just would the token cost be insane, but you'd also get a completely different implementation each time, maybe with different UI and design decisions where the spec left things open.

The alternative is to see the code as the source of truth and LLMs as (extremely sophisticated) editing or refactoring tools. Then by all means, still check in your prompts, but now they are documentation on how a feature was implemented, not the source of truth themselves.

  src/
    md/
      README.md          # index of all md, entry point for agents
      TODO.md            # a list of general TODOs open for this module
      OVERVIEW.md        # a technical overview of this module
      features/FEATURE_1.md       # a set of feature-specific documents
      data/DATAMODEL_1.md         # descriptions of data models in the module
      api/API_1.md                # descriptions of APIs the module provides
      infrastructure/INFRASTRUCTURE_1.md   # descriptions of infrastructure used by the module

  src/
    README.md    # entry point for humans and agents
    TODO.md      # list of general TODOs open for this module
    INFRA.md     # descriptions of infrastructure used by the module
    api/
      README.md  # descriptions of APIs the module provides
    models/
      README.md  # descriptions of data models in this module

It's a SEXP language but could just as easily be markdown. It'd probably be better if it was - I spend a lot of tokens matching parentheses.

Compared to markdown - I lose some human readability but gain a lot in workflow.

For me - it turns the sort of "free form jazz code odyssey" into something with structure. It changes the question to "does the spec match my idea? And does the code match my spec?"

Every so often, I do a "code astrology reset" and try to purge all the tricks I use to keep things functional. I absolutely notice an efficiency hit when I don't have a spec language. Even with the new models - I find this essential to keep things on track.

I think most people on large projects hit a scaling ceiling with markdown specs. They get massive and contradictory very quickly. I think a src/md folder is a good strategy. I try to organize with a "one spec per module" rule. It doesn't always happen that way but I find that helpful.

I think Literate Programming is a good inspiration here. I also think Yegge's beads and Gastown have some really smart things to say in this regard - even if he's a bit tokenmaxxy for my taste.

The interesting part of any program source is what it took to make that program. I view AI-generated code in the same category that I see a binary.

[1] https://blog.tombert.com/Posts/Technical/2026/04-April/Stop-...

Essays of this nature should generally be prefixed with "if you happen to work in exactly the same way that I do..."

Before implementation, context is always cleared to ensure that the model is working only from known context (skills, AGENT.md, etc).

A major benefit of this is that agents can look up prior prompt instructions as needed or when mentioned in the current prompt it's working on.

I recently described my workflow here https://jaisenmathai.com/articles/sojourn-for-ios-was-45-one...

1. https://softengbook.org/articles/deep-modules

IME developers have low tolerance for interrupting flow to track down the docu that is aimed at them, so keep it near and obvious.

One thing to say given some of the comments here: in this approach the stuff in /src/md would be treated as well as stuff in /src/java or whatever: it would need to be maintained as precisely as possible, not just be a stream of development record of what has happened.

I'm also happy to consider that Markdown might not be the right format for this, or that the whole thing is a terrible idea. But I see people saving content in other places (linear, etc.) that drive agents and it just seems like maybe we should keep that stuff close to the generated source.

1. he's describing a manual version of Spec-Kit / OpenSpec / BMAD?

2. Thinks Markdown is... somehow different from normal text or other markup languages?

3. Doesn't seem to understand that LLMs are not deterministic?

i'd be enthused to work this way if changes to `src/*.md` files could be "hot-reloaded" on save, like an iterative LLM watch compiler, and skipping the chat prompts altogether.

If you split code up into multiple packages, moving the docs closer to the code might make sense. The docs for package foo could be in packages/foo/docs.

First, none of this changes the need for a separate `/docs` dir, also checked into git. Wikis are fun, but docs are essential!

Second, I personally think the new paradigm will be putting all of this into tons of new README.md files, which I'm kinda baffled aren't more common deep into dir hierarchies already. That intuitively tells the human authors and the artificial readers that;

A) it toes a similar "for technical people but not necessarily just our dedicated engineering team" line as the root README --more formal than an ephemeral "prompt" and less formal than a user-facing doc,

B) this isn't the AGENT.md file so should remain human-authored only,

C) this is only an overview with a strong preference for brevity & clarity, and

D) this is focused on this specific directory (along w/ the other benefits of locality, as the author extolls already).

Have I cracked the code? Is there a Turing award for inventing the concept of using a tool we already use but just a bit more extensively -- or at least a YC slot?

*P.S.* OP you dropped this: )

> That Markdown should be checked in to /src

I've made a similar argument last June, although I made the point with a sample project rather than manifesto style

https://github.com/TZubiri/keyboard-transpositions-checker

> A Proof of Concept for LLM as source code. I am of the position that the "code" is now target code and not source code. My strongest source for this claim is Stallman's classical definition of Source Code which is the "preferred way to modify the program".

> Upload the conversation with ChatGPT as source code

Although this was pre-agents/markdown, the main way to generate code was just with user-interface conversations with the LLM.

And my position differs in which I argue against uploading the target code, just the source code. Which is an issue if the LLM is not deterministic, but it's workable.

That said I apparently did foreshadow solutions to these issues

> Since we are using a consumer interface, and not an api, (first we have no guarantees that the code will not be used for training, so there's no IP protection, but also), we don't have much reproducibility, robust publishing should use API access and publish exact model with revision as well as seed (and even then we are not guaranteed reproducibility.

I think in general we'll end up rewinding so much, the sloppers are recursing and building a ton of harnessing tools that will need to be trashed because they made a fundamental mistake at the very first steps. Those of us who are a bit more careful will probably only get back to agentic engineering once the slopper activity dies down, it's pretty much career suicide at this point to get into it. Anyone who is barely careful is just waiting it out.