A Tale of Two Claude Code Questions

Two questions about the same tool crossed my path on Stack Overflow recently, and they could not have met more different ends. "How do you index a codebase in Claude Code?" sat at minus eight, closed, down in the company of homework dumps and a request for help cracking a password. Meanwhile an answer to "Is reviewing only by sight a good use of Claude Code to write code?" scored 13, which on today's Stack Overflow is rare air, the league of a beautiful walkthrough of CRC arithmetic and a pile of C++ standardese.

Same tool. Same site. Same community voting. Opposite verdicts. I have read both questions more times than they deserve, because I think the gap between them is not noise, and it is not hostility to AI either. The community is sorting, and once you see the sorting rule, you can use it.

The one that died

Start with the corpse. "How do you index a codebase in Claude Code?" is a reasonable-sounding sentence. It is also, and I say this with sympathy, three distinct mistakes wearing one question mark.

First, it is built on a premise the tool does not have. Claude Code does not build an index of your codebase. It searches on demand, the way you would with grep, and the entire design bets that fast search beats a stale index. Asking how to configure the index is asking how to saddle a car. There is no answer to the question as posed, only a correction of it, and questions that can only be answered by dismantling their premise fare badly everywhere, not just on Stack Overflow.

Second, it is a support question. Not a lesser question, but a different genus: its true answer lives in the vendor's documentation and changes with the vendor's release notes. Stack Overflow has been closing tool-support questions since long before this tool existed, because the site's format assumes an answer worth archiving for a decade, and support answers rot in a quarter. When a question's shelf life is shorter than its review queue, closing it is not gatekeeping. It is inventory management.

Third, and this is the mistake I actually care about, it asks about the tool and not about the work. Which is exactly the trap I wrote about in the essay on chasing tools: the tool is the part you can swap in a day, and the part underneath, the part that took years, is where the durable questions live.

The one that lived

Now the survivor. "Is reviewing only by sight a good use of Claude Code to write code?" has no premise to dismantle, because it does not claim anything about how the tool works. It has no shelf life problem, because it will still be a live question three tools and five versions from now; swap the product name and it was a live question in the era of outsourcing, and of code review generally. And it is not a support question, because no vendor documentation can settle it. It is a judgment question. The only honest answer is experience: someone who reviewed by sight, shipped the result, and found out what that costs.

The answer that scored 13 is worth something precisely because it cannot be generated on demand. A model will happily produce an essay on review practices, and it will be fluent and generic. What the voters rewarded was different: a person staking a position from their own scar tissue, on a question where positions are all there is.

Full disclosure, this one found a soft target in me too. I published an essay on exactly this question about a week before the site's voters lifted that answer into rare air, arguing that reading a diff is real review for one kind of change and a comforting story for another. I do not know who asked the question or who answered it, and I am not claiming the essay caused anything. What I am claiming is smaller and more interesting: the question is in the air. When the same topic surfaces independently on a personal weblog and on Stack Overflow within the same few weeks, that is what a live question looks like from two vantage points.

The sorting rule

Put the two questions side by side and the rule almost states itself.

Ask whether the next release could make your question meaningless. If the vendor ships a feature tomorrow, or renames one, and your question evaporates, it is a support question, and the crowd of strangers on a Q&A site is the wrong audience for it. The right audience is the docs, the changelog, the vendor's own channels, or ten minutes with the tool itself, which will demonstrate its actual behavior faster than a forum can describe it. Nothing shameful in the category. It is just perishable, and archives are for durables.

But if your question would survive the tool's disappearance, if you could cross out the product name, write a competitor's in its place, and still need the answer, then you are not asking about the tool at all. You are asking about the work: what to delegate, what to verify, what a practice costs, where judgment has to stay human. Those questions have no expiry date, no documentation to be redirected to, and no premise to collapse. They are also, not coincidentally, the questions AI answers worst, because the answer is not retrieval. It is testimony.

The community's voting was harsh on one side of that line and generous on the other, and I think the harshness and the generosity are the same signal read twice. A site that archives answers for decades has every reason to reject what rots and hoard what keeps. AI did not create that logic. It just moved the line, by making everything retrievable rot faster.

Asking in the open

If there is advice in this, it is not "avoid asking about AI tools in public". It is: before you ask, find out which of the two questions you are holding. The support question deserves ten minutes with the documentation, and honestly, the model itself will answer it better than a forum will, since describing its own behavior is the one topic on which it is the primary source. The judgment question deserves the opposite treatment. Take it to people, and dress it properly: name the stakes, say what you tried, stake your own tentative position so the testimony has something to push against.

The two questions were asked by two strangers who will never know they were a matched pair. Between them they mapped the whole territory: what not to bring to other people, and what other people are still, irreplaceably, for. Minus eight and plus thirteen is not a contradiction in the community's mood about AI. It is the most precise statement of it I have seen all year.