Troubleshooting
Most problems with Claude, ChatGPT, and Gemini trace back to a few causes: a file the tool cannot read, a conversation grown too long, or instructions that drifted out of view. This page is organised by symptom — find the matching row, read the cause, fix it.
File Upload Failures
| Symptom | Likely cause | Fix |
|---|---|---|
| Upload spins forever or times out | File too large for the tier | Split the document, or upload only the relevant pages/sheets |
| "Unsupported file type" | Format the platform cannot read | Convert to PDF, plain text, CSV, or PNG/JPG first |
| PDF uploads but the AI "can't read it" | Scanned image with no text layer | Run OCR (Acrobat, Preview), or upload page screenshots |
| Spreadsheet uploads but numbers look wrong | Merged cells or multiple tabs | Export a single clean tab as CSV with one header row |
When a file will not cooperate, paste the content directly instead:
I couldn't upload the file, so here is the raw content below.
----- BEGIN DOCUMENT -----
[paste text, CSV rows, or table here]
----- END DOCUMENT -----
Now [summarise / extract the totals / find the dates].
For multi-file work, upload to a persistent surface. In Claude, open the Projects panel in the left sidebar, create a Project, and add files to its knowledge. In ChatGPT, keep files in one conversation where Code Interpreter is active. See the file handling lesson.
Context Window Exhaustion
The context window is the AI's short-term memory. When a conversation grows past it, the oldest messages silently fall out of view — so early instructions, files, or decisions get "forgotten."
| Symptom | Likely cause | Fix |
|---|---|---|
| AI forgets a rule you set earlier | Instruction scrolled out of the window | Re-state the rule, or move it to Custom Instructions / a Project |
| It re-asks for a file you uploaded | File dropped out of context | Re-attach it, or store it in a Project's knowledge |
| Replies turn vague or repetitive | Window saturated with old turns | Start a fresh conversation; paste a short summary |
| Long document gets partial answers | Document exceeds what fits at once | Process it in labelled chunks, then merge |
Carry context into a fresh chat with a handoff prompt:
Continuing from a previous conversation. Resume from this state:
GOAL: [one sentence]
DECISIONS SO FAR: [bullet list]
CURRENT TASK: [what to do next]
CONSTRAINTS: [rules to keep following]
Confirm you have this, then continue.
Durable rules belong somewhere persistent, not mid-chat: use ChatGPT's Custom Instructions (profile, then Settings), a Claude Project's instructions field, or Gemini's saved info. The power-user rules cover conversation hygiene in depth.
Inconsistent or Unreliable Responses
| Symptom | Likely cause | Fix |
|---|---|---|
| Same prompt gives different answers | Normal model variability | Lower ambiguity: specify format, length, examples |
| Confident but wrong facts | The model is guessing | Ask it to cite sources or enable browsing; verify yourself |
| Ignores part of your instructions | Too many asks in one message | Number the requirements; ask it to confirm each |
| Output format drifts over a long chat | Earlier formatting fell out of context | Re-paste the format spec each time, or use a template |
| Answer is shallow | Prompt invited a shallow answer | Ask for reasoning first, then a conclusion |
A structured prompt removes most inconsistency:
ROLE: [who the AI should act as]
TASK: [the single, specific job]
FORMAT: [table / bullets / 200 words]
MUST INCLUDE: [non-negotiable elements]
If anything is unclear, ask before answering.
For factual work, force verification rather than trust the first reply: "List your sources," or "Flag any claim you are not confident about." More live in the prompting patterns reference.
Feature Access and "Missing" Features
Features differ by platform, subscription, and app. A capability missing in one place often exists in another.
| Symptom | Likely cause | Fix |
|---|---|---|
| Code Interpreter / data analysis missing | Gated to a paid tier | Check your plan; switch to the model variant with tools |
| Artifacts / Canvas panel not appearing | Wrong mode or an older client | Update the app; ask explicitly for an Artifact or Canvas |
| Web browsing returns stale info | Browsing not enabled for that model | Pick a model/mode with live search turned on |
| GPTs or the store not visible | Free tier or region limits | Custom GPTs need a paid ChatGPT plan |
| Gemini absent from Docs/Gmail | Workspace integration off | Confirm the right account and that Gemini is on for it |
Before assuming a feature is broken, confirm you are signed into the correct account, your plan includes the feature, and your app is up to date. Request features by their proper names — Projects, Artifacts, Canvas, Code Interpreter, GPTs, Gems — so the tool surfaces the right one.
Subscription and Account Confusion
| Symptom | Likely cause | Fix |
|---|---|---|
| Paid features missing despite paying | Signed into the wrong account | Check the email under your profile; sign out and in |
| Usage limit hit quickly | Heavy model has tighter caps | Switch to a lighter model for routine tasks; wait for reset |
| Billed but tier looks like Free | Payment pending or wrong workspace | Check billing under Settings; confirm personal vs team |
| Team member can't see a feature | Admin policy disabled it | Ask your workspace admin to enable it |
When limits bite, match the model to the job: reserve the most capable model for hard work, and use a faster, cheaper one for drafts. The platform cheat sheet lists which tiers unlock which features.
Slow or Sluggish Long Conversations
| Symptom | Likely cause | Fix |
|---|---|---|
| Chat scrolls and types slowly | Huge transcript straining the app | Start a new conversation; archive the old one |
| Responses arrive slowly | Heavy model or peak demand | Switch to a faster model; retry off-peak |
| Browser tab becomes unresponsive | Memory exhausted by a long session | Close other tabs, refresh, or use the desktop app |
The single best habit is one conversation per task. When a thread has done its job, summarise the outcome, start fresh, and paste the summary forward using the handoff template above. Short, focused threads stay faster and more accurate than one endless thread.