Limitations
What Claude and Claude Code are bad at today, and the workflows that compensate.
Why this page exists
Every limitation below is a known property of current frontier models, not a bug in Claude specifically. The point of listing them is operational: if you know where the agent is weak, you can design workflows that catch the failure before it ships.
The limitations that bite most often
Confidently wrong
The agent will state incorrect facts — API signatures, flag names, library behavior — with the same tone it uses for correct ones. It is not lying; it does not know it is wrong. Defense: verify any factual claim against the source before acting on it, especially version-specific ones (command flags, schema fields).
Loses detail over long context
In a very long session the agent can drop or conflate earlier constraints. A rule you
set in message 3 may be forgotten by message 40. Defense: put standing rules in
CLAUDE.md, not in chat. Keep sessions scoped; /clear or
/compact when the thread has outlived its purpose.
Bad at precise counting and exact arithmetic
Ask “how many TODO comments are in this directory?” and you may get a number that is
close but wrong. Defense: for anything that must be exact, have the agent run a real
command (grep -c, a test, a linter) and read its output rather than estimating.
Suggests plausible-but-nonexistent APIs
The agent may invent a function signature or a CLI flag that would be reasonable if it existed. Defense: for any unfamiliar API, open the docs. This is why this site exists — to point you at the source rather than the agent’s guess.
Over-eager to please
If you imply an answer in your question (“should I use pattern X here?”), the agent tends to agree. Defense: ask open questions (“what pattern fits this case, and why?”) rather than leading ones.
No real understanding of your runtime
The agent can read your code, but it does not know your production data, your traffic, or your users. Defense: for anything with real consequences (migrations, deploys, schema changes), treat the agent’s output as a draft for a human, not a decision.
What this means for your workflow
The reliable pattern is the same one this site’s project plan is built on:
- Draft with the agent. It is fast at first drafts.
- Verify against the source. Especially facts.
- A human approves. Never auto-ship.
This is not a workaround for a model that will soon be fixed — it is the correct workflow for any tool that is reliable-but-not-certain. See Security for the do/don’t list that follows from these limitations.
Pitfalls
- Treating confidence as correctness. Tone is not evidence.
- Letting the agent’s draft become the spec. The draft is input to your judgment, not a replacement for it.
- Skipping review because “the agent wrote tests.” Generated tests can assert the wrong thing and pass. Read them.