Why does Claude Code or Cursor get my product's API wrong?
A coding assistant works from the docs it can fetch, or from older copies it learned before, so stale, hard-to-fetch or ambiguous docs lead to wrong code. Make your current docs reachable, in plain text, with versioned examples and a clear statement of what changed. A short llms.txt can point agents to the right pages, though it is only a proposal.
Free to read, no signup. Nova checks how an assistant describes your API and where the docs fall short.
What is documented about fetching docs
No coding assistant publishes how it chooses docs. The crawler documentation and the llms.txt proposal are the closest published guidance.
Claude fetches pages when asked
Anthropic says Claude-User fetches websites when a user asks a question that needs them, and that blocking it can reduce visibility there. Your docs need to be reachable.
Does Anthropic crawl data from the web?llms.txt is a proposal, mostly for docs
The llms.txt site describes a proposed Markdown file that links agents to LLM-friendly pages, mostly used for software documentation, and says OpenAI, Anthropic and Google publish their own.
The /llms.txt fileOpenAI's user-triggered fetcher
OpenAI says ChatGPT-User visits pages when a user asks a question, and that robots rules may not apply to user-initiated visits.
Overview of OpenAI crawlersFive steps to fix how assistants see your API
Do it by hand with these steps, or hand the whole thing to Nova.
Ask the assistants
Ask coding assistants to write code for your API and note wrong or outdated calls.
Nova: asks assistants about your API and logs mistakes.
Match mistakes to docs
For each mistake, find the doc page, if any, that states the correct way.
Nova: matches each mistake to the right page.
Make current docs reachable
Confirm docs are public, not behind logins or heavy scripts, and not blocked from bots.
Nova: checks that your docs can be fetched.
Write for agents too
Add plain-text, versioned examples, a changelog and clear deprecation notes. Consider a short llms.txt.
Nova: drafts doc fixes and an llms.txt.
Recheck
Re-ask the assistants in a few weeks and compare.
Nova: re-checks and reports.
Short on time? Nova, your AI brand visibility expert, runs this Maneuver on your brand and tracks the result.
Take it to NovaMistakes, causes and fixes
| Assistant says | Correct | Cause | Fix |
|---|---|---|---|
| Call [OLD ENDPOINT] | Call [NEW ENDPOINT] | Old docs still live | Redirect and note deprecation |
| Use [OLD PARAM] | Use [NEW PARAM] | No changelog | Add a changelog |
| No such feature | Feature exists | Docs behind login | Make docs public |
Nova then re-asks the assistants, so you see whether the mistakes persist.
Take this Maneuver into your own AI agent
Get docs-discoverability.md, a ready-to-use skill file with a mistake log, a docs reachability check and a doc fix list. Drop it into Claude, Cursor or any agent that reads SKILL.md files.
Skill file for Claude, Cursor and other agents
- An assistant mistake log
- A docs reachability check
- A versioned-examples checklist
Questions, answered.
Relevant reads
All articlesLet Nova check how assistants see your API
Nova opens with this Maneuver loaded. Nothing ships without your approval.