GenAI Field GuideWhere it fails
Common pitfalls
Almost every bad outcome with these tools comes from one of a dozen habits. Here they are, with the fix.
| Pitfall | What happens | Fix |
|---|---|---|
| Official material in a personal app | Data leaves the accredited boundary | Never a personal account for official material; your accredited platform, within your command's guidance. Where your text goes |
| Treating it like a search engine | Answers from stale training data, stated confidently | Attach the source or use search; know its knowledge cutoff from the model card |
| No source given | Plausible invention | "Answer only from the attached; say if it isn't there" |
| Trusting the citations | Fabricated references that look real | Open every citation before you use it |
| One giant prompt | The fourth thing you asked for is done badly | One job per prompt; check between steps |
| A 200-turn chat | Early instructions forgotten after compaction | New chat per task; durable material in a project |
| Numbers from a language model | Arithmetic and counts that are subtly wrong | Calculator, spreadsheet, or code you read |
| Wrong tool for the shape of the job | Fighting the product instead of working | Which tool for which job |
| Accepting an agent's "done" | Silent failures reported as success | Give it a test for done; read the transcript |
| Powerful tools plus untrusted content | Prompt injection turns a document into instructions | Narrow tools, human gate before actions |
| Not reading the output | Fluent text goes out with an error in paragraph three | Read it as the recipient before you sign |
| Chasing the newest model | Slower, no more accurate for routine work | Fit, source, thinking level, verify. Choosing a model |