Best Practices · Optional
What Not to Do
The common mistakes that turn a capable assistant into a liability, and the simple habits that avoid each one.
Most of this guide tells you what to do. This page does the opposite, because the fastest way to get good results is often to stop doing a few specific things. None of these mistakes is dramatic, and that is exactly why they are dangerous: an invented retention figure does not announce itself, it simply flows into an IC paper. They are quiet, easy habits that slowly erode the quality and safety of your work, and each one has a simple fix. Read this once, recognize yourself in a couple of them (everyone does), and adjust.
Do not trust output you have not checked
This is the big one, and it earns its own page in Verify Everything. The shorthand: fluent does not mean correct. A confident, well-formatted figure can be wrong, and you cannot tell from how it reads. Never let an unverified number, quote, or claim reach a deliverable. If you do nothing else differently, do this.
Do not upload before you have made the confidentiality call
Dragging a document in to start working is the moment a restricted file ends up in a personal account or a workspace it was never cleared for. Run the checks in Your Account & Confidential Data first, every time.
Do not give a vague brief and expect a sharp answer
“Have a look at this and tell me what you think” is not a brief, it is a shrug. Claude will respond to it with an equally unfocused answer, and you will blame the tool for a problem you created. The fix is the four-part shape from Anatomy of a Good Prompt: say who Claude is helping, what you want done, how you want it back, and what to watch for. A precise question is most of a good answer.
The most common form of this is attaching a document with no instruction at all: a sixty-page CIM dropped in with the word “thoughts?”. The file is rich, but Claude has to guess what you actually care about, and it usually guesses wrong, returning a generic overview that misses the point. The file is the same either way; the instruction is what turns a shrug into a usable screen. Attach the document, then say exactly what to do with it.
Do not leave the format to chance
If you do not say what shape you want the answer in, Claude picks one, and it is rarely the shape you needed. You then spend time rewriting a wall of prose into the table, the four headings, or the five bullets you wanted from the start. Decide the format up front and say so: “a table”, “four headings”, “five short bullets”. It costs a few words and saves the rewrite.
Do not mix several deals in one conversation
Pasting bits of two different targets into the same chat to “save time” does the opposite. The context blurs, and Claude starts conflating one company’s figures with another’s. Keep each engagement in its own Project, as covered in Setting Up a Deal Project. One deal, one space.
Do not restart when you should refine
Re-prompting from scratch every time an answer is slightly off wastes the context you have already built. If a draft is mostly right, correct the part that is wrong with a follow-up. This is the refine-in-place pattern from Anatomy of a Good Prompt, and it is faster and produces better results than starting over.
Do not ask Claude to make the decision
Claude is a brilliant first-draft writer and a fast synthesiser. It is not an investment committee. Asking it to decide whether to pursue a deal, or treating its conclusion as the answer rather than an input, hands away the one thing that is irreducibly yours. Use it to assemble the case, surface the risks, and argue both sides. Then you decide.
Do not assume more words means more rigor
A long, fluent answer can feel thorough and be hollow. Length is not a proxy for correctness or for judgment. A three-line answer that traces every claim to source is worth more than three pages of confident prose you have not checked. Read for substance, not volume.
The mistakes and their fixes, at a glance
Common mistakes and the habit that fixes each
| Mistake | The fix |
|---|---|
| Trusting unverified output | Check every figure, quote, and claim against the source |
| Uploading before the confidentiality call | Run the upload checks first, every time |
| Vague, one-line briefs | Use the four-part shape: role, task, format, scrutiny |
| Dropping in a file with no instruction | Attach the document, then say exactly what to do with it |
| Leaving the output format to chance | Name the shape up front: a table, four headings, five bullets |
| Mixing several deals in one chat | One Project per engagement, kept separate |
| Restarting instead of refining | Correct the wrong part with a follow-up |
| Letting Claude make the decision | Use it for the draft; keep the judgment |
| Equating length with rigor | Read for substance and traceable claims, not volume |
Each fix is small. Together they are the difference between a liability and a capable partner.
What to take from this
Almost every failure with Claude traces back to one of these habits, and almost every fix is a small change in routine rather than a new skill. Verify before you trust, brief precisely, keep deals separate, refine rather than restart, and keep the judgment with you. For the positive version of all this on a single page, see the Cheat Sheet.