Common tasks

Everyday jobs that an agent does well, such as small edits, docs, and release chores. The theme of this chapter: AI is good at things you do not want to spend time on. None of these tasks are hard; they are just friction. An agent turns them into a one-line request you can fire off while you do something else.

All the example prompts below are real prompts from real sessions; where the result is public, the prompt links to the PR it produced.

Small edits¶

Config tweaks, version bumps, adopting a new tool -- these are perfect fire-and-forget tasks:

Let’s bump the upper limit on cmake to 4.4

— scikit-build-sample-projects#79

Let’s bump the actions to the latest versions in this PR. npx actionsup or something like that might help.

Notice the second prompt mentions a tool that might help, without being sure of the details. That’s fine; the agent will figure it out.

Merge conflicts and rebases¶

Repetitive conflict resolution does not need a strong model or your hands. Often one word is enough:

rebase

With AGENTS.md set up (see AGENTS.md), the agent tests the result and fixes fallout too. It scales up naturally:

Let’s rebase all the PRs made by me now that CI works again.

And it can fold in related work while it is there:

Rebase and update the changelog (try nox -s make_changelog)

Rebase, cuda’s been moved to extras

— cuda-histogram#69

Conversion tasks¶

CI provider migrations, config format changes (tox.ini → tox.toml), docs system moves, language ports, citation formats:

Let’s convert the gitbook to mystmd

— CLI11#1389

Let’s convert this example to nanobind

Let’s convert this to the new recipe format. There’s an automated tool somewhere for that, can be run with pixi I think.

— histoprint-feedstock#21

Fixing lints¶

Enable a lint with no autofix and let the agent iterate over the report. More lint rules make agent output better overall, so this is a win-win: the agent cleans up the existing violations, and the rule then keeps future agent (and human) code honest.

Can we also lint with prek on PRs?

— SIMPLE-Py#23

Let’s update for ruff v0.16.0 - this can now be handled by ruff itself for markdown, anyway. The guide might need a few updates for 0.16 too.

— cookie#833

Better docs¶

Agents are good at converting code into readable text that pairs with the code. They match existing style, move text inline, and fill missing details. A review pass is as easy as:

Review the docs

They are also good at finding drift between code and docs, and at making sure it does not come back:

In the docs, cp314-pyodide_wasm32 is missing. Are there any others missing? Is there a way to keep this from happening again?

— cibuildwheel#2947

Docs still refer to dynamic_metadata.plugins.<plugin>, we removed the .plugins part.

— dynamic-metadata#88

One-off scripts¶

Plots, data analysis, Dockerfiles from shell history, quick tooling -- output you use once and do not need to review as production code:

Let’s make a script that makes a plot of the minimum scikit-build-core versions set in these pyproject.toml’s. The script can use PEP 723 for dependencies (packaging, plotting tools), and run with uv run.

Small web work fits here too: pages, JavaScript/TypeScript when you are rusty. You can describe visuals in plain words:

The github and documentation buttons don’t have the frosted glass effect on the main page, they should have it too (tiles do already)

— scikit-build.github.io#10

Release chores¶

Draft release notes or a changelog from the git log between two tags; the agent mimics your existing style:

What’s new since last release?

Fill out the changelog since last release

Let’s prepare the changelog and version number for a new patch release.

— CLI11#1410

Merged. Let’s make the gh release. Use the changelog to fill the description (as markdown).

— plumbum v2.0.2

Reviving old work¶

Ask it to bring an outdated PR up to date with the current codebase, even one that has been dead for years:

Let’s rebase #4272 and see if we can make it work.

— pybind11#4272, opened in 2022 and merged in 2026

I want to reopen #1142 and rebase it on main (it was pointing at a branch that got merged).

— nox#1142, reopened and merged

Tasks you never had time for¶

Small cleanups, dependency swaps, restructuring; the backlog items that were hard to justify time for are now just a sentence and a little time in the background:

Let’s remove the “needs-changelog” label infrastructure, it’s not really needed anymore, and not kept up to date.

— boost-histogram#1164

Also things you wouldn’t have done, because they were too much work, like testing downstream projects before and after a change:

What downstream projects need updates for this, from the ones I work on: packaging, cibuildwheel, cmake, scikit-build-core, ...

Low-friction mixing¶

You can mix your own writing with agent work at any granularity. Leave TODO comments while writing, then ask the AI to find and fix them. This works for prose too:

Here’s a draft message, help me fill in the TODOs: ...

— the filled-in message became the description of boostorg/histogram#429

You keep the structure and the voice; the agent fills in the mechanical parts.