> Basically, try to keep the content in these files to an absolute minimum
focusing on information that cannot be inferred/discovered, one-line code
patterns, disallowed behaviours, links to guides.

Absolutely. And as usual.. PRs are welcome :)

On Mon, Mar 2, 2026 at 10:40 AM Nathan Hadfield <[email protected]>
wrote:

> Hello,
>
> On the topic of CLAUDE/AGENTS.md files, there was a rather interesting
> paper published recently about their effectiveness.
>
> https://arxiv.org/abs/2602.11988
>
> The TD;LR is
>
>
>   *
> LLM-generated context files reduce success rates (0.5-2%) while increasing
> inference cost by 20-23%
>   *
> Developer-written files help slightly (+4%), but verbose content that
> duplicates existing docs is pure cost
>   *
> Codebase overviews don't improve navigation: agents find relevant files in
> the same number of steps regardless
>
> Basically, try to keep the content in these files to an absolute minimum
> focusing on information that cannot be inferred/discovered, one-line code
> patterns, disallowed behaviours, links to guides.
>
> Hope this helps.
>
> Cheers,
>
> Nathan
>
> From: Jarek Potiuk <[email protected]>
> Date: Monday, 2 March 2026 at 09:29
> To: [email protected] <[email protected]>
> Subject: Re: [DISCUSS] Active approach to fighting with AI slop (while
> keeping maintainers in the driving seat)
>
> This Message Is From an External Sender
> This message came from outside your organization.
>
>
> Also: I am thinking of more tools like that - especially one that will
> allow us to auto-triage issues and use an LLM to speed up issue
> classification for provider releases (once suggested by Shahar I think) and
> many more things.
>
> The quality of good models is amazing. I am literally stunned by what
> Claude Code can do today - I tried it few months ago and the difference is
> night and day. I literally entirely Claude-Coded the whole thing
> without writing a single line of code myself.
>
> And since we have at the very least 6 months of free Claude Code Max for
> maintainers of big OSS projects
>
> https://urldefense.com/v3/__https://claude.com/contact-sales/claude-for-oss__;!!Ci6f514n9QsL8ck!mS8qTovb9go2kfJwcUGGry6yWpOOcdvB2IXJYYHcOEam-B2gTQQ_dcYm19lzIlgAKCiUragw0XqPXOJZ$
> as of 3 days (liteally day
> after I paid for my first month)!!!) - Airflow definitely qualifies, so all
> core maintainers can get it regardless if their employees already pay them
> for it.  So if you have not done it yet - apply :D.
>
> J.
>
> On Mon, Mar 2, 2026 at 10:22 AM Jarek Potiuk <[email protected]> wrote:
>
> > > maybe we should use the new LLMOperator form common.ai as an option
> > (hehe)!
> > Just joking, of course.
> >
> > Crossed my mind :D
> >
> > On Mon, Mar 2, 2026 at 10:20 AM Pavankumar Gopidesu <
> > [email protected]> wrote:
> >
> >> This is really cool, Jarek. Thanks for sharing. A tool like this is
> >> definitely necessary given the current volume of AI slope and PRs being
> >> submitted without proper context.
> >>
> >> maybe we should use the new LLMOperator form common.ai as an option
> >> (hehe)!
> >> Just joking, of course.
> >>
> >> Regards,
> >> Pavan
> >>
> >> On Mon, Mar 2, 2026 at 9:17 AM Jarek Potiuk <[email protected]> wrote:
> >>
> >> > > I think that we could later automate at least the dry-run execution
> of
> >> > the
> >> > script, along with Slack notification for highly-suspected issues/PRs.
> >> > Then, it would be easier for maintainers to react fast when needed.
> >> >
> >> > Yes. I would like to run it manually—ideally with several volunteer
> >> > maintainers - for a while to see how it works, improve and iterate and
> >> > possibly add more quality gates. When we have more confidence we could
> >> run
> >> > it automatically for some parts or even the whole process eventually
> >> > (especially for high-confidence/sensitive stuff), keeping the
> sensitive
> >> > parts with Human-In-The-Loop.
> >> >
> >> > But also (and this is my hope) - similarly to `breeze ci upgrade` it
> >> might
> >> > turn out that the process is so efficient and "nice" to follow that we
> >> > could continue trigger it manually, regularly, perhaps with a
> rotational
> >> > maintainer handling the triage. I think comments and actions coming
> >> from a
> >> > human maintainer have more value than those from a bot—even if the
> human
> >> > action is merely confirming what an automated system or LLM proposed.
> >> >
> >> > J.
> >> >
> >> >
> >> > On Mon, Mar 2, 2026 at 10:04 AM Shahar Epstein <[email protected]>
> >> wrote:
> >> >
> >> > > Amazing stuff Jarek!
> >> > > I think that we could later automate at least the dry-run execution
> of
> >> > the
> >> > > script, along with Slack notification for highly-suspected
> issues/PRs.
> >> > > Then, it would be easier for maintainers to react fast when needed.
> >> > >
> >> > > Looking forward for new AI-based features in breeze in particular,
> and
> >> > > Airflow in general :)
> >> > >
> >> > >
> >> > > Shahar
> >> > >
> >> > >
> >> > > On Sat, Feb 28, 2026, 04:59 Jarek Potiuk <[email protected]> wrote:
> >> > >
> >> > > > Hello everyone,
> >> > > >
> >> > > > While preparing for consensus on the assignment policy, I created
> PR
> >> > > >
> https://urldefense.com/v3/__https://github.com/apache/airflow/pull/62585__;!!Ci6f514n9QsL8ck!mS8qTovb9go2kfJwcUGGry6yWpOOcdvB2IXJYYHcOEam-B2gTQQ_dcYm19lzIlgAKCiUragw0bHQEs-s$.
> This PR adds a new
> >> > command
> >> > > > to
> >> > > > Breeze, `breeze issues unassign`, which unassigns anyone who is
> not
> >> a
> >> > > > committer or collaborator.
> >> > > >
> >> > > > I want this to be the first of several Breeze commands I plan to
> >> add to
> >> > > > help manage the AI overhead and burden on maintainers.
> >> > > >
> >> > > > I got inspired bu Hugo van Kamerade's (my friend, Python release
> >> > manager)
> >> > > > tool
> https://urldefense.com/v3/__https://hugovk.dev/blog/2026/gh-triage/__;!!Ci6f514n9QsL8ck!mS8qTovb9go2kfJwcUGGry6yWpOOcdvB2IXJYYHcOEam-B2gTQQ_dcYm19lzIlgAKCiUragw0TLnNyxs$.
> He added the `gh`
> >> > plugin
> >> > > > that helps him manage spam coming to Python. I hope we can have
> very
> >> > > > similar set of commands and regular process of performing cleanup
> >> with
> >> > > the
> >> > > > issues/prs we are getting.
> >> > > >
> >> > > > BTW. I am using Claude Code to add those commands (so this is a
> bit
> >> > like
> >> > > > using AI to fight AI slop). But in a smart way.
> >> > > >
> >> > > > In our case we have `breeze` that we are already using for `ci
> >> upgrade`
> >> > > by
> >> > > > maintainers and I see no reason why we could not use our own CLI
> to
> >> > make
> >> > > us
> >> > > > far more efficient with assessing and quickly and efficiently
> >> > processing
> >> > > > incoming spam.
> >> > > >
> >> > > > Starting with AGENTS.md that describes what we expect (and
> instructs
> >> > > agents
> >> > > > to make good PRs) and changing our assignment process - I think we
> >> > should
> >> > > > proceed to implement step-by-step handling of the incoming
> traffic:
> >> > > >
> >> > > > a) Quickly assess how well PRs implement our expectations, point
> out
> >> > > > problems, and close them
> >> > > >
> >> > > > b) automatically telling the collaborators what is wrong with
> their
> >> PRs
> >> > > if
> >> > > > they are incomplete (for example when tests are failing, or when
> >> they
> >> > > need
> >> > > > a rebase)
> >> > > >
> >> > > > c) automatically responding to issues that they are incomplete and
> >> need
> >> > > > more information
> >> > > >
> >> > > > d) Allow filtering by area (so that maintainers focusing on a
> >> > particular
> >> > > > area can periodically review only the areas they are intereste
> >> > > > e) all that with some AI assistance (I plan to imlpement
> integration
> >> > with
> >> > > > some modern AI LLMs so that it is seamless for those maintainers
> who
> >> > > > already use some of those (including Cloud Code, GH Copilot
> >> > (maintainers
> >> > > > can apply for free access there), Codex and any models someone
> >> prefers
> >> > -
> >> > > > including local models).
> >> > > >
> >> > > > f) all that with maintainer in the driver's seat—we won't do those
> >> > things
> >> > > > fully automatically - but we will get reviewable action proposal
> in
> >> > bulk
> >> > > > that the maintainer will be able to accept, modify or reject.
> >> > > >
> >> > > > .... more...
> >> > > >
> >> > > > All that will be open to contribution and I will be happy to
> leading
> >> > > > introduction and disseminating those CLI options between
> >> maintainers to
> >> > > > make sure those get incorporated in our daily work - relieving
> some
> >> of
> >> > > the
> >> > > > burden we are all experiencing and sharing it between people.
> >> > > >
> >> > > > I think this is a viable approach to address our current burden
> >> > > > proactively, rather than waiting for others to act.
> >> > > >
> >> > > > This is also somewhat experimental since we haven't seen it done
> >> > before,
> >> > > so
> >> > > > suggestions, comments, ideas and PRs that could help us become
> more
> >> > > > efficient and better maintainers are most welcome.
> >> > > >
> >> > > > Let me know what you think.
> >> > > >
> >> > > > J.
> >> > > >
> >> > >
> >> >
> >>
> >
>
>

Reply via email to