
The word assistant hides three products. There is an assistant you think with, one you produce with, and one you hand work to. They look similar — a text box and a reply — and they are used very differently.
The three in the directory that make the distinction concrete: Pi, ChatGPT, and Claude Code.
The short answer
Pick the mode first. A thinking partner for decisions, a production tool for artefacts, a delegation agent for tasks you can verify. Using the wrong mode produces the classic complaint: 'it gives generic answers' or 'it changes things I did not ask about'.
Thinking mode
Pi is built for conversation rather than output: it asks follow-ups and keeps a warm register. If you have a dilemma, this mode is more useful than a tool that immediately produces a plan.
The failure of thinking mode is when it is used to produce: a reflective assistant is not going to give you a correct API signature.
Production mode
ChatGPT is a toolbox: files, tools, search, and a deliverable at the end. It is the right mode when the question is 'make this' — a draft, a summary, a table.
The failure of production mode is over-production: the model writes more than was asked, and the user accepts a longer artefact rather than a better one.
Delegation mode
Claude Code takes a task and acts on a repository. It is the mode where the output is a change, not a text. That requires a different review: a diff, not a read.
The failure of delegation mode is scope: the agent does what it can rather than what you meant, and the result looks complete.
The comparison
| Mode | Input | Output | Review |
|---|---|---|---|
| Think | A situation | A better question | Your own thinking |
| Produce | A brief | An artefact | Read it |
| Delegate | A task | A change | Diff and tests |
The review column is how you know which mode you are in. If you are reading for correctness, you are producing; if you are reading a diff, you are delegating.
Why people pick the wrong mode
Two reasons. The first is habit: whatever tool is open gets every question. The second is cost: the strongest model is used for thinking, and the cheapest for producing, when it should often be the reverse — thinking benefits from quality, production benefits from iteration.
A worked week
Monday: a decision about direction. Think with the assistant, do not ask for a plan. Tuesday: a report to write. Produce with the tool that has file access. Wednesday: a backlog of small tickets. Delegate to the agent, review diffs. Three modes, three tools, one workflow.
Frequently asked questions
Can one assistant do all three? Yes, badly. The mode mismatch produces the specific complaints people have about each category.
Which mode is cheapest? Thinking and producing can use cheaper models; delegation needs a strong model because the check is a diff. See choosing a daily-driver assistant.
Is Pi worth a separate tool? If you have ever wanted to talk a decision through, yes. If you only produce, no. See general assistants.
Last checked: 16 November 2026.
Related: Pi vs ChatGPT, choosing a daily-driver assistant, and general assistants.


