2026-01-06 ensure your work is a meaningful contribution

A developer reflects on the importance of delivering proven, working code rather than merely generating large patches with AI assistance. The post emphasizes that agent-generated code is cheap and that meaningful contributions require testing, honest naming, and self-consistent abstractions. It warns that sloppy PRs cost more than they save and that human accountability remains essential.

| meaningful contribution | | | Based on embarassing feedback on a bad PR for AXM-10598, and written from my beliefs; with only a little bit of copy pasting, because to be human is to copy from a shitload of places until it becomes your own. | | | A contribution is not code, it's proven working code | | | "your job is to deliver code you have proven to work." — simon willison https://simonwillison.net/2025/Dec/18/code-proven-to-work/ | | | agent-generated code is cheap. anyone can prompt an LLM to produce a thousand-line patch. that's not valuable. what's valuable is contributing code that demonstrably works, has been tested, and doesn't shift burden to reviewers. | | | what qualifies as meaningful | | | 1. proven correctness | | | if you haven't tested it, it doesn't work. | | | if it happens to work, that's luck. | | | to actually prove something works, we: | | | - you tested manually, and saw it work. | | | - you wrote an automated test for what you did manually. | | | - you saw this automated test fail when you reverted you change. | | | - you tested the edges, you explicitly defined what happens outside the happy path. | | | 2. self-consistent abstractions | | | you created a mental model, great, | | | does it make sense when you look closer? | | | does it fit with the wider mental model? | | | naming something VersionedStructuredRequestWithOptions and then passing unversioned requests through it is confusing. names are contracts. | | | i don't know how to prove that something makes sense, but a good start is: | | | - you can explain it in plain english; from start to finish, and each part in isolation. | | | - you can explicitly articulate the expected inputs and outputs, and what happens if one of those is unexpected. | | | what doesn't qualify | | | slop indicators | | | - PR descriptions that read like "summarize this for me" prompts | | | - missing tests | | | - contradictions in abstractions, both against themselves and against the wider context. | | | - names that lie about what they contain, if you need a comment or "x but it's actually y or x+y", its lying. | | | the speed trap | | | speed without quality is negative value. a sloppy PR costs more reviewer and maintenance time than it saves. the "i'll fix it later" debt compounds negativelly, the "i made it well" value compounds positivelly. | | | we lost track of the goal. we sacrificed quality in pursuit of speed, and for what? | | | the review standard | | | ask yourself before submitting: | | | 1. have i seen this work? not "does the code look right"—have i actually run it? | | | 2. do the types tell the truth? or am i lying to the compiler and hoping reviewers don't notice? | | | 3. is the naming honest? would someone reading this in six months be confused? | | | 4. did i test the edges? what happens when usage deviates from the happy path? when it's the WORST path? | | | 5. would i be confident to walk colleagues through my changes? | | | if the answer to any of these is "no" or "i'm not sure," the contribution isn't ready. | | | agent-assisted work requires more scrutiny, not less | | | agents make it easy to produce large volumes of code quickly. this doesn't reduce your responsibility—it increases it. the code still has your name on it. | | | one pass from an agent is rarely enough. read it over. improve it. run it. test it. do this multiple times if needed. | | | the accountability loop | | | a computer can never be held accountable. that's your job as the human. almost anyone can generate a thousand-line patch. what's valuable is proving it works. | | | next time you submit a PR, include the evidence. |