Skip to content

Build1 publisher3 min readPublished

A failed docs answer can open the pull request that fixes the page

One developer-docs maintainer routes his assistant's failed answers into a job that drafts a change and files a pull request. The trigger reads the assistant's own reply, so a confident wrong answer still needs a human to flag it.

The Engineer · Build desk

Illustration accompanying A failed docs answer can open the pull request that fixes the page

What happened

  • Two inputs feed the same system: automatic detection when the assistant's reply looks like "not found" or only partial context, and human feedback through tags, escalation and a feedback form.
  • Mintlify survey findings cited in the post say most teams now have AI agents drafting documentation updates, while only a small group lets those agents publish without a human in the middle.
  • The same findings, as summarised, say a large share of teams take a week or longer for a product change to reach the docs and only about a quarter get it there the same day.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

  • constraint Catching the confidently wrong page depends on a reader noticing and tagging it, so the blind spot grows as agent readers take over traffic from the humans who would have complained.
  • cost Drafting is automated and merging is not, so the recurring cost of this design is reviewer attention, and the pull request queue is where the loop stalls.
  • exposure An agent that reads a stale page carries the stale instruction into whatever it builds next, so a late doc update lands in generated code rather than in a support thread.
  • decision Teams already running a docs bot have to decide whether its miss log is wired to the documentation repo or merely stored somewhere.

Automatic gap detection fires on the shape of the assistant's own reply: a "not found" answer, or one that came back with only partial context [10]. The second input is the reader, through helpful and not-helpful tags, a "needs a doc update" mark, a bug report, or free-form feedback into Slack [11]. The author of the dev.to post separates what each one catches: automatic detection finds problems left out, and human feedback catches "the answer was there, but it was not good enough" [19]. Either signal becomes a data point, and the data point can open an automated documentation job that drafts a change and files a pull request; a human reviews it and decides what to merge [9].

The gap in that arrangement is the confidently wrong page. "Sometimes the assistant answers confidently, and the reader still knows something is off," the author wrote [12]. Only the human path catches that case, and the premise of the whole project is that the fastest-growing class of reader will not notice. An agent, the author wrote, "does not usually pause and ask whether the page is correct; instead, it treats the docs as instructions and keeps going from that incorrect information" [3]. By that description, the input that scales with agent traffic is the reply-shape detector, and the detector cannot see a wrong answer that is well formed. The post does not disclose what share of the site's traffic comes from agents, or how many pull requests the job has produced [18].

Adoption starts well before the automation. You need an assistant on the docs site that readers actually use, which the author says is the case every day [14]. Then capture of every miss, then a job with write access to the docs repo, then someone to read its output [9]. "If the assistant cannot answer, and that failure disappears into a log nobody reads, you only built a nicer search box; that's it," the author wrote [13].

The post cites Mintlify survey findings. "The interesting part is not the exact percentages," the author wrote [7]. As reported, a large share of teams take a week or longer for a product change to show up in docs, and only about a quarter say it usually lands the same day [5]. Roughly three in four teams therefore take longer than a day [17]. Most teams in the same survey say agents now draft their documentation updates, and only a small group lets those agents publish without a human in the middle [6].

The author's test for whether a docs bot is doing anything is whether a failed answer turns into work in the documentation repo [15]. It is a test any team with an assistant can run against its own miss log. The process it replaces ran through people: an outdated page opened a ticket to a support engineer in Slack, the product team passed the message on, and the author fixed and republished the page [2].

What to watch

  • Part 2 is promised on the assistant itself, and Part 3 on packaging docs for coding agents through indexes and skills.
  • How many of the drafted pull requests survive human review, a figure the post does not give.
  • Whether the Mintlify survey's actual percentages get published alongside a loop like this one.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories