TL;DR: I propose adding a devlog/ directory to the PyTorch repo for short markdown posts — the same technical write-ups we already do in Workplace, but checked into the repo where OSS users can read them and AI agents index them automatically. Contributors writing important posts are highly encouraged to push them to the repo. In a case study, this cut AI response time from 9 minutes (mediocre) to 38 seconds (correct, with citations). PR: https://github.com/pytorch/pytorch/pull/178721
DevLogs: One Practice That Helps AI Assistants, OSS Users, and Your Future Self
A thought that kept hitting me whenever I wrote a intenral post i or saw a great one written by someone else — was: how can I share this with OSS users with minimal effort? Even for internal users, these posts get maybe days, weeks at most, before they get lost in the Workplace feed. When you need them, you can’t easily find them.
These posts are often very well-written summaries, directional proposals, or deep technical dives. DevDiscussion, PyTorch blog posts, GitHub issues… never felt like the right place.
And in the AI era, AI coding assistants answer questions usually by searching code — but code only tells you what exists, not why it was built, what it replaced, or whether it worked. Even finding the right code is hard — a general question like “how does PyTorch handle data-dependent shapes?” doesn’t map neatly to any filename or function. The knowledge that makes an answer actually useful lives in those same Workplace posts that agents can’t see nor can OSS users.
What do I like is to have?
I want plat form that is :
-
Easy to publish. A PR review at most. If you write a Workplace post, you can make it a devlog entry with minimal effort.
-
Easy to access by OSS users and AI agents. It lives with the code, so OSS users see it when they browse the repo and AI agents index it automatically. No special tooling, no separate platform.
-
Easy to navigate/index. Entries are organized by topic and prefixed with dates. If you want to know what happened in Dynamo over the past three years, you open a single Dynamo file that has a list of all Dynamo posts. AI agents also index this easily and use it as a first search step — scanning titles first, not full post contents.
The Proposal
https://github.com/pytorch/pytorch/pull/178721
A devlog/ directory in the repo with short markdown posts. Each component has a top-level directory with a table of all posts.
All sub-posts are in subdirectories, example: https://github.com/…/2026-03-25-unbacked-perf-parity.md
CLAUDE.md and Devmate instructions:
-
For every question, list files under
devlog/compile/first -
Prefer newer entries (date-prefixed filenames)
-
For “how/why” questions: read matching posts before searching code
-
For code tasks: only read a post if the title is clearly relevant
-
Cite the date and title of the used post in the results
-
Inform users that the answer is from the devlog and ask if they want a code search or if the answer is sufficient
AI Quality and Response Time: A Case Study
Aditya Venkataraman suggested this doing this to test.
I asked: “How did PyTorch solve the data-dependent error problem?”
Without devlog — P2256431852 9 minutes.139.5k tokens Mediocre, incomplete answer mentioning related tools/APIs without connecting them.
With devlog — P2256394622 38 seconds. Correct answer, higher quality.
Identified : “The fix: Explicit Unbacked Semantics”
Cited sources: “Per the 2025-10-29 Slaying Framework Data-Dependent Errors…”
End *“This answer is based on devlog entries. Would you like me to search the code to verify or supplement?”
thoughts?*