Build 1 publisher2 min readPublished September 20, 2026 at 4:11 AM PDT

A Claude routine wrote a README into a pull request from Anthropic's own infrastructure

A routine is a prompt, a repository and a trigger saved in Claude's web console and executed on Anthropic's machines. Switching one on means installing a GitHub app on the repository you choose.

The Engineer · Build desk

Illustration accompanying A Claude routine wrote a README into a pull request from Anthropic's own infrastructure

What happened

Anthropic's docs describe routines as putting Claude Code on autopilot, defining jobs that run on a schedule, trigger on API calls, or react to GitHub events from cloud infrastructure.
In a dev.to walkthrough, a pull request opened from GitHub's web interface came back a short time later carrying a new section titled "Add comprehensive Readme".
The GitHub event trigger ships with four event types out of the box: pull request, PR merged, release published and issue opened.
The first run can report that the GitHub app is not installed, at which point the user follows a link, grants permissions, selects a repository and installs the integration.
Notifications default to push messages in the Claude mobile and desktop apps, email to the address on the account, and Slack if Slack has been connected.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

decision Teams picking routines over Actions move the reviewable artifact from a YAML file in the repository to a prompt in a vendor console, so a change to what fires on every pull request no longer arrives as a diff.
exposure The install grants a GitHub app permissions on the repositories you select, so anyone who can edit the routine's prompt can change what happens to those repositories.
capability Unattended scheduled work no longer needs a hosted runner or a machine left switched on, because Anthropic runs the job while the developer's own Claude instance is off.
constraint The record of a run lives in the routine's panel in Claude, not in the CI system the rest of the team already watches, so debugging a failed automation starts somewhere new.

When a pull request opens, GitHub hands the event to Anthropic. The routine executes there, against whichever repositories you connected to it, and the pull request is what comes back changed ^[2].

The definition of that automation is four fields on the routines page of the Claude web interface: a name, a prompt, a repository and a trigger ^[3]^[4]. The prompt in the dev.to walkthrough told Claude what to do with the README ^[4]. The commit that set it off was a one-line comment on the shell script that starts python's http.server for a browser game ^[11].

Three trigger types exist ^[5]. Schedule runs at the time you set, on Anthropic's infrastructure, including when your own instance of Claude is not running ^[6]. The GitHub event trigger is the one the walkthrough used. The API trigger issues a URL and a token and fires on a POST, so another service or an external orchestrator can call it ^[7].

The author wrote that he started by asking himself "could I get rid of actions entirely?" ^[13] On this evidence the answer turns on what the job tolerates. The unit of configuration here is a prompt, where an Actions workflow would pin a step, and the task under test was writing documentation that a human then reads in the pull request. A release gate or a test matrix has less slack for a differently worded result each run.

The post does not cover pricing, concurrency limits or retry behaviour. Those are the numbers a platform team would need before moving anything a build depends on off its own runners. The path exists and takes about five clicks to wire up, which is a lower adoption cost than any self-hosted runner I have stood up.

What to watch

Whether the out-of-the-box GitHub event catalogue grows past four types, which decides how much of an Actions workflow can actually leave.
Whether Anthropic ships org-scoped routines with shared ownership, instead of one account holding the prompt and the notification address.
Whether run logs become exportable in a form an existing CI dashboard can ingest.

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Anthropic's official Claude docs, as quoted in the dev.to post, say: "Put Claude Code on autopilot. Define routines that run on a schedule, trigger on API calls, or react to GitHub events from cloud infrastructure."
ReportedSupportedSource: Anthropic documentation, quoted by dev.toView cited source
[2]
Routines run on Anthropic's infrastructure, can be connected to one or more repositories, and can respond to API requests or to GitHub events such as pull requests.
ReportedSupportedView cited source
[3]
Routines are saved configurations (prompts) inside Claude Code, created by logging into the Claude Code web interface and going to the routines page.
ReportedSupportedView cited source
[4]
Creating a routine requires giving it a name and taking three actions: add a prompt, select a repository, and select a trigger; the author's prompt gave instructions on what to do with the README.
ReportedSupportedView cited source
[5]
There are three types of triggers: schedule, GitHub event, and API.
ReportedSupportedView cited source
[7]
The API trigger gives a URL and a token so the routine can be called via a POST request, which the author describes as flexible enough to embed in other code or an external orchestrator.
ReportedSupportedView cited source
[8]
The author selected a GitHub pull request as his trigger; other out-of-the-box GitHub event types are PR merged, Release published and Issue opened.
ReportedSupportedView cited source
[9]
The default notification options for a routine are push notifications to the Claude mobile and desktop apps, email to the address on the account, and Slack, which requires Slack to be connected.
ReportedSupportedView cited source
[10]
On the first run the user may see a note that the GitHub app is not installed; the user clicks the link to install the app, grants the appropriate permissions, selects a repository, clicks install, and is redirected to the installed connectors page in the Claude web browser showing the GitHub integration installed.
ReportedSupportedView cited source
[11]
The repository had no README present at all; the author added a single comment to the shell script that starts python's inbuilt http.server to serve his HTML space invaders game, committed it to a new branch, and opened the pull request from GitHub's web-based interface to prove nothing was running on his laptop.
ReportedSupportedView cited source
[12]
After a short amount of time the pull request was updated with information from the routine, including a new section saying "Add comprehensive Readme".
ReportedSupportedView cited source
[13]
The author wrote that he thought to himself "could I get rid of actions entirely?" and set about to find out.
ReportedSupportedSource: dev.to post authorView cited source
[14]
Once the workflow is created, the routine's details appear in a right-hand side panel along with any runs that have occurred.
ReportedSupportedView cited source
[d1]
The GitHub event trigger ships with four out-of-the-box event types: pull request, PR merged, release published and issue opened.
DerivedSupported1 source— sign in to open
[6]
Scheduled routines are set for particular times and run on Anthropic's infrastructure, and they run when the user's personal instance of Claude is not running.
ReportedInsufficientView cited source

Sources

Publishers with included, body-backed reporting in this cluster.

dev.to
1 article · September 20, 2026
Using claude routines with github

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Try

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Everything needed to copy this is in the dev.to post. A routine takes a name, a prompt, a repository and a trigger, and the trigger choices are schedule, GitHub event or API. The GitHub event list is short: pull request, PR merged, release published, issue opened. The API trigger hands back a URL, a token and a pre-filled curl example, and firing it returns a session id. The first run stops until you install the GitHub app and grant it access to a repository you pick. And the author never exercised the schedule trigger, so the cron case here rests on Anthropic's documentation.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence42

0100

One walkthrough, one developer

One developer's walkthrough is the whole basis, written as he went. The steps are specific and the stumbles are reported: the first run stalled on the missing GitHub app, and the API call fired a routine that then did nothing because no pull request was open. The central result, a README committed to a pull request, sits in screenshots a reader cannot open, and nobody has repeated the test.

Adoption10

0100

A single personal repository

The whole of it is one routine on the author's space invaders repository, fired by one pull request, and the README it wrote was that repository's first. The GitHub app went onto the single repository he selected. No team, employer or second user appears anywhere in the account.

Hype gap+12

−100 understated0 aligned+100 overstated

Framing outruns the test

The post opens on one question and closes on something narrower. "Could I get rid of actions entirely?" opens the post; a committed README closes it, and no Actions workflow was retired or even compared. Pulling the other way, the author published the run that did nothing and said so plainly.

Incentives25

0100

No stake, vendor's words

The author sells nothing and declares no relationship with Anthropic. The frame is the vendor's, though: the post opens on the documentation's "Put Claude Code on autopilot" line and then exercises the feature that line advertises, on a hobby project where a bad commit breaks a space invaders game.

Confidence45

0100

Solid on mechanics, thin beyond

The click-by-click detail is easy to believe, and an author who publishes a run that accomplished nothing is not curating much. Confidence falls away on what he did not touch: what the GitHub app's permissions actually grant, and whether a scheduled routine behaves as the documentation says.

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Entities

build
Running Cline in CI means switching off the approval gate it ships with
1 publisher
build
The weights never moved: what 6,852 Claude Code sessions say about where regressions live
1 publisher
build
Claude Code's new default is a confession: the approval prompt was never a control
1 publisher
build
Copilot's credit meter moves the cost decision into the model dropdown
1 publisher

What happened

Why it matters

What to watch

Claim ledger

Sources

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Related stories

Claim ledger

Sources

Let Clarity write the post for you.

Perspective Coverage

Reality

Topics and entities

Topics

Entities

Related stories