agentleFS
pandas-devpandas-dev/pandas/.agents/skills/pandas-review/SKILL.md

pandas-review

Skill51k starsChanged 2 days agoBSD-3-ClauseView on GitHub

Self-review your own pandas PR (or your local branch before opening one) the way a pandas maintainer would. Prints findings and a verdict in the terminal; never posts to GitHub or edits files.

Adopted by

Steady: 0 new projects took it in the last 4 weeks, none the 4 weeks before.thin data

Week of Sep 20: 0 new projectsWeek of Sep 27: 0 new projectsWeek of Oct 4: 0 new projectsWeek of Oct 11: 0 new projectsWeek of Oct 18: 0 new projectsWeek of Oct 25: 0 new projectsWeek of Nov 1: 0 new projectsWeek of Nov 8: 0 new projects
Sep 2026Oct 2026Nov 2026
New projects taking a version of this file each week, over the 8 weeks our latest run read: only 0 in all, so this is a hint, not a trend.

Who's talking about pandas-dev/pandas

The biggest 5 of 77 mentions linking to the repository that our crawl found. Each line is their own headline.

See all 77

Talked about on Stack Overflow, YouTube, GitHub, Hacker News and the web (77 times).

On Stack Overflow · 28

23 more on Stack Overflow

On YouTube · 23

18 more on YouTube

On GitHub · 14

9 more on GitHub

On Hacker News · 11

6 more on Hacker News

On the web · 1

  • In this repositoryOne other file it gives its agents.

The file

We split this skill into the rules it gives an agent

11 rules. About 0 are common, 7 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read all of them in full, never truncated: diff.patch, any untracked files it lists, and context.md (the PR description, reviews, inline …

    line 31
  2. Very specific:

    When the code under review is the PR head on GitHub, its changed files are saved under head/.

    line 38
  3. Distinctive:

    If it reports the branch is behind main, say so.

    line 41
  4. Very specific:

    Read AGENTS.md at the repo root.

    line 50
  5. Distinctive:

    Run code when the local build is the code under review: Run the PR's new and changed tests, the existing test files covering the code it …

    line 54
  6. Very specific:

    Maintainer position: If a maintainer has said the PR shouldn't proceed, the verdict is CLOSE.

    line 63
  7. Distinctive:

    Don't nit comment length otherwise.

    line 89
  8. Distinctive:

    Every finding ends with a concrete fix.

    line 112
  9. Distinctive:

    Keep it concise: findings, not a checklist of what you inspected.

    line 133
  10. Distinctive:

    If you disagree with a finding, say why on the thread in your own words rather than changing code to satisfy this review.

    line 157
  11. Distinctive:

    Don't modify files or git state.

    line 166

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Review a pandas pull request the way a pandas maintainer would, and report to its author.

Target: $ARGUMENTS

The target is a PR number or URL, a branch name, or nothing (including when the line above reads literally $ARGUMENTS) to review the current branch with its uncommitted changes.

Review as a maintainer who did not write this code, even if you did. Verify what the PR description, commit messages, and code comments claim instead of taking them on trust. Report only: don't edit files, and don't fix anything during the review. The author decides what to change afterwards.

1. Gather the change and its thread

From the repo root, run

python .agents/skills/pandas-review/scripts/fetch_context.py <target>

with no argument for the current branch, and owner:branch for a branch in a fork. It prints the PR it found (if any), where the code under review is, and the files it wrote. Read all of them in full, never truncated: diff.patch, any untracked files it lists, and context.md (the PR description, reviews, inline comments, conversation, the issues the PR closes or its description references with their reactions, and failing CI jobs with log excerpts).

  • If the script fails, stop and report its error. A failed fetch is unknown, not empty; never review as if the thread were silent.
  • When the code under review is the PR head on GitHub, its changed files are saved under head/. Read any other file at the head with gh api -X GET "repos/pandas-dev/pandas/contents/<path>?ref=<head sha>" -H "Accept: application/vnd.github.raw".
  • If it reports the branch is behind main, say so. CI tests the PR merged into main, so a recent change on main can break it there even when it passes locally.
  • Review verdicts matter most: a maintainer's approval, requested changes, or "I don't think this is worth doing" usually lives there, not inline.
  • The thread outranks this file. Where a maintainer on the thread has asked for something that conflicts with the guidance below, follow the maintainer.

2. Investigate

Read AGENTS.md at the repo root. Consult doc/source/development/contributing_codebase.rst when a question about tests, typing, or backward compatibility comes up. Use Grep/Glob/Read on the surrounding code to check the change against how the codebase already does things.

Run code when the local build is the code under review. Run the PR's new and changed tests, the existing test files covering the code it touches, and every repro you report, and show real output. If pandas fails to import or build, say so and review from reading alone; don't try to repair the environment.

3. What to look for

In priority order:

  1. Maintainer position. If a maintainer has said the PR shouldn't proceed, the verdict is CLOSE; if they want a different approach, WRONG APPROACH. Skip to section 4 and report only that. If they've said it's blocked on a decision, the verdict is BLOCKED, but review the rest to give the verdict it would get once unblocked. A core maintainer's 👍 reaction on the linked issue counts as buy-in. A /take with no activity for several days doesn't block anyone.
  2. Unaddressed review comments. For each reviewer request in the thread, check whether the current diff addresses it or the author replied with a reason. List any that are neither. Don't draft replies: pandas asks that authors speak for themselves on the thread (see doc/source/development/contributing.rst). If the author has already explained on the thread why one of your findings doesn't apply, say so and don't count it toward the verdict, unless the explanation is factually wrong (then say what is wrong with it).
  3. Correctness. Bugs in the diff, and above all regressions: inputs that worked on main and break with this PR.
    • A fix that covers some code paths but leaves sibling paths unfixed is not a defect. At most mention it as an optional follow-up. Only regressions and wrong results on the paths the PR claims to fix drive NEEDS WORK.
    • A failing CI job the diff causes is a correctness finding; quote the log line.
  4. Necessity, simplicity, brevity. Flag only concrete excess you can point at:
    • code, parameters, or tests the fix doesn't need (a guard no input reaches, with the reason; a refactor riding along; a test duplicating an existing one);
    • a new helper duplicating existing code (name it at file:line), a one-caller abstraction that inlines with no loss, or a materially smaller equally-correct approach (sketch it);
    • the same rationale pasted at several sites, or comments narrating code the PR removed. Don't nit comment length otherwise. Equal-size alternative designs and naming taste are not findings.
  5. Conventions that commonly come up in review:
    • A whatsnew entry describes the user-visible change, not the implementation, and names only public API. A pure refactor gets no entry.
    • Tests: module-level functions or an existing class are both fine; so is a new class with a reasonable scope.

Do not report:

  • things that are fine ("tests look good", "no whatsnew needed"); silence means no concern;
  • PR process or hygiene: template checkboxes, labels, merge readiness;
  • CI failures the diff didn't cause (infrastructure, flaky tests, a break already on main);
  • drift between the PR description and the diff; review the diff;
  • predictions of what a pre-commit hook or CI will catch (the author should run pre-commit separately).

4. Output

Open with one or two sentences on what the PR does, then the findings, most important first.

Anchor each localized finding at path:line on its own line, using the line number at the head of the code under review (not a hunk offset). Use a range when it spans lines. For diffuse findings (wrong approach, missing coverage) use prose; don't invent an anchor. Every finding ends with a concrete fix.

pandas/core/indexes/multi.py:1487
`codes` is reused after the early `continue`, so the vectorized path sees the previous
level's array. Fix: move the assignment above the branch.

One repro per distinct failure mode: runnable Python, imports included, actual next to expected. Different inputs that fail the same way share one repro. Say in one line whether you executed it; never present predicted output as observed.

import pandas as pd

idx = pd.MultiIndex.from_product([["a", "b"], [1, 2]])
print(idx.get_locs([["a"], [1, 2]]))
# actual:   [0]
# expected: [0, 1]

Keep it concise: findings, not a checklist of what you inspected.

End with the verdict, the last thing in the output, in a fenced box:

┏━━━━━━━━━━━━━━━━━━━━━━━┓
┃  VERDICT: NEEDS WORK  ┃
┗━━━━━━━━━━━━━━━━━━━━━━━┛
VerdictMeans
LGTMReady for maintainer review as-is.
NITS ONLYNothing blocking; the findings are optional.
NEEDS WORKReal defects; the approach is sound.
WRONG APPROACHCan't be patched into correctness as written.
BLOCKEDWaiting on a maintainer decision or another PR, not on the author. Add the verdict it would get once unblocked: BLOCKED → NITS ONLY.
CLOSEUnlikely to be merged in any form; say why.

For the author

  • NITS ONLY clears the bar for review, but LGTM gets reviewed sooner. Fix the nits you agree with and rerun once; then stop, since further reruns surface new optional nits, not blockers.
  • If you disagree with a finding, say why on the thread in your own words rather than changing code to satisfy this review.
  • A green verdict doesn't replace your own review of the change, which pandas requires, and your PR description still needs the tool, model, and effort you used (see the automated contributions policy in contributing.rst).

Hard rules

  • Everything stays in the terminal. Never post, comment, review, react, or push on GitHub.
  • Don't modify files or git state.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.