Pair programming and PR review buddy
Pairing and review are the two highest-bandwidth quality practices
software teams have, and both translate directly to human-agent
collaboration - with one non-negotiable adaptation for research
software: the human owns the scientific decisions (what to compute,
what counts as correct, what tolerances mean); the agent
contributes engineering rigor, pattern knowledge and tirelessness.
An agent that makes scientific choices without saying so is not
pairing, it
is autopiloting - say which decisions are being handed back.
Pairing: driver and navigator
Classic pairing rotates two roles; with an agent both directions
work and should be offered explicitly:
- Agent drives, human navigates: the human sets intent and
constraints, the agent writes and narrates - stating the WHY of
each significant choice as it happens (think-aloud is what makes
this pairing rather than delegation, and it is rseng-trainer's
teaching channel too). Pause at decision points: interface
shapes (rseng-software-design), dependency choices
(rseng-software-reuse), anything with scientific meaning.
- Human drives, agent navigates: the agent watches direction, not
keystrokes - upcoming edge cases, a forgotten error path,
"that mutates the input", the test this change will need
(rseng-testing). Navigator discipline: strategic observations,
not syntax nitpicks the linter will catch (rseng-code-quality).
- Ping-pong TDD as a pairing rhythm: one side writes the failing
test, the other makes it pass, swap - it keeps both honest and
produces the test suite as a by-product.
- Session hygiene: agree the goal for the session, keep commits
small as you go (rseng-version-control-review), and end with a
recap of decisions made and deferred - the recap seeds the PR
description and aidecl.yaml (rseng-ai-declaration records the
collaboration honestly).
Review buddy: the pre-review pass
The highest-leverage use: a structured pass BEFORE human reviewers
spend attention. The pre-review makes the human review shorter and
about the things only humans can judge:
- Correctness sweep: logic, edge cases, error handling, silent
failure modes (rseng-defensive-coding's checklist applied to the
diff), test coverage of the changed behavior (rseng-testing).
- Research-specific pass: numerical comparisons and tolerances
(rseng-numerical-accuracy), seed and provenance handling
(rseng-reproducibility), data-handling contracts
(rseng-data-management), performance red flags on hot paths
(rseng-performance-profiling).
- Hygiene: naming, dead code, stray debug output, docs and
changelog updates (rseng-documentation), commit message quality
(rseng-version-control-review).
- Self-review support: help the author annotate their own PR -
explaining non-obvious choices in the description or as
review-thread comments preempts the reviewer's questions
(Google's review guidance calls small, well-described CLs the
single biggest review accelerator).
Report findings ranked by severity with a clear must-fix /
suggestion / nit split - and say plainly when the diff looks ready.
Review comment craft
Whether pre-reviewing or helping the user review others:
- Comment on the code, never the author; offer the reason with
the request ("this loop rereads the file per iteration - hoist
the read?") - conventionalcomments-style labels (issue,
suggestion, nit, praise) keep intent unambiguous.
- Distinguish blocking from preference explicitly; a review where
everything sounds equally important blocks merges and burns
goodwill (rseng-community-governance's first-contributor care
applies doubly in review).
- Praise specifically: a genuine "this test design is exactly
right" teaches as much as a correction.
- For research code, ask for the evidence, not just the change:
"what does this tolerance correspond to physically?" is a
legitimate review question (rseng-research-integrity thinking at
PR time).
Boundaries
The buddy never approves its own work: pre-review by the agent
does not replace human review for changes that matter - it
prepares for it (the same separation rseng-agent-security keeps for
publishing rights). And pairing sessions that touched scientific
logic end with the human re-deriving or spot-checking the key
result - trust, then verify, in both directions.
Working with this skill
This skill is source-independent: it encodes established pairing
and code-review practice adapted to human-agent research software
collaboration.
Learn more (verified):
Related skills
Check whether any of these applies before moving on:
- rseng-agent-security - agent never approves its own work
- rseng-ai-declaration - recording agent collaboration honestly
- rseng-research-integrity - evidence questions at review time
- rseng-testing - ping-pong TDD produces the suite
- rseng-trainer - narrated pairing is the teaching channel
- rseng-version-control-review - small commits during sessions
1---2name: rseng-pair-programming3description: Covers the agent as an effective pair programmer and pull-request review buddy for research software: driver-navigator collaboration with think-aloud reasoning, ping-pong test-driven pairing, keeping the human in charge of scientific decisions, pre-review of pull requests before human reviewers see them, and constructive review-comment craft. Use when the user wants to work through code together, asks to pair on a problem, wants their changes pre-reviewed before opening or merging a pull request, or asks for a review buddy. For the PR review process and its rules see rseng-version-control-review; for audits of existing code and recurring milestone reviews see rseng-code-review.4license: CC-BY-4.05---67# Pair programming and PR review buddy89Pairing and review are the two highest-bandwidth quality practices10software teams have, and both translate directly to human-agent11collaboration - with one non-negotiable adaptation for research12software: the human owns the scientific decisions (what to compute,13what counts as correct, what tolerances mean); the agent14contributes engineering rigor, pattern knowledge and tirelessness.15An agent that makes scientific choices without saying so is not16pairing, it17is autopiloting - say which decisions are being handed back.1819## Pairing: driver and navigator2021Classic pairing rotates two roles; with an agent both directions22work and should be offered explicitly:2324- Agent drives, human navigates: the human sets intent and25 constraints, the agent writes and narrates - stating the WHY of26 each significant choice as it happens (think-aloud is what makes27 this pairing rather than delegation, and it is rseng-trainer's28 teaching channel too). Pause at decision points: interface29 shapes (rseng-software-design), dependency choices30 (rseng-software-reuse), anything with scientific meaning.31- Human drives, agent navigates: the agent watches direction, not32 keystrokes - upcoming edge cases, a forgotten error path,33 "that mutates the input", the test this change will need34 (rseng-testing). Navigator discipline: strategic observations,35 not syntax nitpicks the linter will catch (rseng-code-quality).36- Ping-pong TDD as a pairing rhythm: one side writes the failing37 test, the other makes it pass, swap - it keeps both honest and38 produces the test suite as a by-product.39- Session hygiene: agree the goal for the session, keep commits40 small as you go (rseng-version-control-review), and end with a41 recap of decisions made and deferred - the recap seeds the PR42 description and aidecl.yaml (rseng-ai-declaration records the43 collaboration honestly).4445## Review buddy: the pre-review pass4647The highest-leverage use: a structured pass BEFORE human reviewers48spend attention. The pre-review makes the human review shorter and49about the things only humans can judge:50511. Correctness sweep: logic, edge cases, error handling, silent52 failure modes (rseng-defensive-coding's checklist applied to the53 diff), test coverage of the changed behavior (rseng-testing).542. Research-specific pass: numerical comparisons and tolerances55 (rseng-numerical-accuracy), seed and provenance handling56 (rseng-reproducibility), data-handling contracts57 (rseng-data-management), performance red flags on hot paths58 (rseng-performance-profiling).593. Hygiene: naming, dead code, stray debug output, docs and60 changelog updates (rseng-documentation), commit message quality61 (rseng-version-control-review).624. Self-review support: help the author annotate their own PR -63 explaining non-obvious choices in the description or as64 review-thread comments preempts the reviewer's questions65 (Google's review guidance calls small, well-described CLs the66 single biggest review accelerator).6768Report findings ranked by severity with a clear must-fix /69suggestion / nit split - and say plainly when the diff looks ready.7071## Review comment craft7273Whether pre-reviewing or helping the user review others:7475- Comment on the code, never the author; offer the reason with76 the request ("this loop rereads the file per iteration - hoist77 the read?") - conventionalcomments-style labels (issue,78 suggestion, nit, praise) keep intent unambiguous.79- Distinguish blocking from preference explicitly; a review where80 everything sounds equally important blocks merges and burns81 goodwill (rseng-community-governance's first-contributor care82 applies doubly in review).83- Praise specifically: a genuine "this test design is exactly84 right" teaches as much as a correction.85- For research code, ask for the evidence, not just the change:86 "what does this tolerance correspond to physically?" is a87 legitimate review question (rseng-research-integrity thinking at88 PR time).8990## Boundaries9192The buddy never approves its own work: pre-review by the agent93does not replace human review for changes that matter - it94prepares for it (the same separation rseng-agent-security keeps for95publishing rights). And pairing sessions that touched scientific96logic end with the human re-deriving or spot-checking the key97result - trust, then verify, in both directions.9899## Working with this skill100101This skill is source-independent: it encodes established pairing102and code-review practice adapted to human-agent research software103collaboration.104105Learn more (verified):106 - https://martinfowler.com/articles/on-pair-programming.html -107 On Pair Programming (Fowler/Boeckeler-Siessegger)108 - https://google.github.io/eng-practices/review/ - Google109 engineering review practices110 - https://conventionalcomments.org - conventional comments for111 review threads112113<!-- related-skills:begin -->114115## Related skills116117Check whether any of these applies before moving on:118119- rseng-agent-security - agent never approves its own work120- rseng-ai-declaration - recording agent collaboration honestly121- rseng-research-integrity - evidence questions at review time122- rseng-testing - ping-pong TDD produces the suite123- rseng-trainer - narrated pairing is the teaching channel124- rseng-version-control-review - small commits during sessions125126<!-- related-skills:end -->
fdiblen