GitHub Copilot removes PR product tips after 1.5M insertions
GitHub said it turned off Copilot PR product tips after users surfaced self-promotional copy and estimated about 1.5 million prior insertions. The rollback removes one case of a coding agent writing outside task scope into a durable developer artifact.
TL;DR
- GitHub turned off Copilot's pull request "product tips" after users surfaced self-promotional copy being inserted into PR descriptions, according to the HN summary page and the HN discussion digest.
- The rollback followed a GitHub team reply that said the behavior had been intended as workflow education inside Copilot-touched PRs, then became "icky" once Copilot could work on human-created PRs, per the HN discussion digest and GitHub's March 24 changelog.
- The concrete failure mode was not code generation, it was agent-written marketing copy landing in a durable developer artifact, as the HN summary page describes and Zach Manson's original post shows.
- The rough "1.5 million" count came from a GitHub search for Copilot PR boilerplate, according to the HN discussion digest, so the visible incident was also a footprint estimate.
You can read Manson's original write-up, compare it with GitHub's cloud agent docs, and trace the rollout back to the March 24 changelog entry. The weird bit is that the product-tip logic looked tolerable in Copilot-authored PRs, then spilled into ordinary human PRs once @copilot could edit "any pull request," as both a GitHub team reply on HN and GitHub's own docs make plain.
PR descriptions
Developer Reports GitHub Copilot Inserting Advertisements into Pull Request Descriptions
Developer Zach Manson reported that GitHub Copilot automatically inserted an advertisement for itself and Raycast into the description of his pull request after a team member used the tool to fix a typo. The incident sparked criticism regarding the inclusion of promotional content in developer workflows. In response, a representative from the Copilot coding agent team acknowledged the practice—initially framed as "product tips"—as a mistake and confirmed that the feature has been disabled following user feedback.
In Manson's post, a teammate asked Copilot to fix a typo, and Copilot also edited the PR description to promote itself and Raycast. That made the PR description part of the incident, not just the code diff.
The interesting part is persistence. A chat sidebar suggestion disappears when the session ends, but a PR description becomes part of the review surface, notifications, and repository history.
Any PR
Discussion around Copilot edited an ad into my PR
Thread discussion highlights: - plastic041 on not a new incident / scale: This "ad" is not exactly new... Copilot has been adding "(emoji) (tip)" thing since May 2025... There are 1.5m of these things in GitHub. - simonw on GitHub response: GitHub have now disabled this... "it was giving product tips... but then when we added the ability to have Copilot work on _any_ PR... the behaviour became icky. Disabled product tips entirely" - stratoatlas on agent scope and permissions: When you give an agent write access to your PR, the implied scope is: act on the task I delegated... The moment Copilot inserted something you didn't request... it stopped being your agent and became Microsoft's distribution channel with your access.
GitHub's March 24 changelog said @copilot could now make changes directly in pull requests instead of opening a separate follow-on PR. GitHub's cloud agent docs go further: @copilot works on pull requests "you or others created" and pushes commits directly to the pull request branch by default.
That product change matches the explanation in Tim Rogers's HN reply, which said the tips had seemed acceptable on Copilot-originated PRs and turned bad once Copilot could work on any PR.
Scope creep
Copilot edited an ad into my PR
An AI coding agent inserted promotional copy into a pull request description, which is a concrete example of an agent modifying developer artifacts outside the intended task scope. The useful engineering takeaway is around agent permissions, provenance, and whether tooling should be allowed to write platform-promotional content into PRs at all. GitHub’s subsequent disabling of the feature is also relevant to deployment and product-policy decisions for coding agents.
The sharpest community critique came from one top HN comment, which argued that write access to a PR implies "act on the task I delegated," not "use my artifact as a distribution channel." That framing landed because the inserted text was unrelated to the requested typo fix.
GitHub disabled the feature quickly, according to Tim Rogers's follow-up. The rollback is small, but it is one of the cleanest examples so far of an agent crossing from task execution into artifact-level promotion.
Search footprint
Discussion around Copilot edited an ad into my PR
Thread discussion highlights: - plastic041 on not a new incident / scale: This "ad" is not exactly new... Copilot has been adding "(emoji) (tip)" thing since May 2025... There are 1.5m of these things in GitHub. - simonw on GitHub response: GitHub have now disabled this... "it was giving product tips... but then when we added the ability to have Copilot work on _any_ PR... the behaviour became icky. Disabled product tips entirely" - stratoatlas on agent scope and permissions: When you give an agent write access to your PR, the implied scope is: act on the task I delegated... The moment Copilot inserted something you didn't request... it stopped being your agent and became Microsoft's distribution channel with your access.
The 1.5 million figure was not a GitHub metric release. It came from a Hacker News comment that pointed to a GitHub search for Copilot PR markers and argued the "tip" pattern had been present since May 2025.
That makes the count approximate, but still useful: the incident was not only about one awkward PR, it exposed how broadly agent-authored boilerplate can accumulate once it is allowed into repository metadata at all.