Why the default output is always average
Ask any assistant to "create a mind map for changing careers" and you get CV, networking, skills, interviews, job search. Every time, from every model.
That is not a flaw in the prompt wording. It is the models doing exactly what they do: producing what is most typical about the subject. Typical is the opposite of useful here, because the branches that decide whether your career change works are the ones specific to your situation — and by construction those are not typical.
So a good prompt is not a more polite request. It is a set of constraints that push the output away from the average.
The four constraints that matter
Split by uncertainty, not topic. The single highest-leverage instruction. Topic splits (health, money, work) are easy to generate and tell you nothing about what will fail. Ask instead for areas that could each fail for separate reasons.
Forbid balance. Models produce symmetrical trees — five branches, four children each. Real subjects are lumpy. Say explicitly that one area should get more depth if it deserves it.
Demand gaps be named. Left alone, a model fills silence with plausible content. Told to flag what it does not know, it will.
Fix the output format. Vague requests get prose. Ask for CSV with named columns and you get something importable.
A prompt that works
Break down [subject] into 4–6 areas that could each fail for separate reasons. Do not split by topic or activity.
For each area, state what would have to be true for it to succeed.
Do not balance the tree. If one area needs more detail than the others, give it more and leave the rest shallow.
Where you are guessing about my situation, say so explicitly instead of filling it in.
Output as CSV with columns: id, parent_id, title, description.
Then, always, one follow-up:
Which of these am I most likely to be wrong about, and what would tell me early?
That question produces more useful editing than any amount of "add more detail." It is also the one people skip.
Give it your actual situation
The largest single improvement is not prompt phrasing. It is context.
Paste the real material — the job description, the syllabus, the brief, the notes — and ask for structure derived from that rather than from general knowledge. The difference between "plan a career change" and "here is the role I want and here is my current CV, plan the gap" is the difference between generic and yours.
Reviewing what comes back
Generated structure arrives sounding complete. Run it through five questions before accepting any of it:
- What is missing that I know matters? The gaps are where the model was averaging.
- Which branch is here because it is common rather than because it applies to me? Delete it.
- What does this assume about my situation? Assumptions are invisible until you go looking.
- Which branch could fail and take three others with it? That is your first Focus, and models almost never identify it.
- What would I have put here that it did not? Your additions are the whole value.
If the review changes nothing, you have not reviewed it. A first draft from something that does not know you is never right.
The thing AI genuinely cannot do
It can propose structure. It cannot set priority.
Priority is not a prediction problem — it is a values problem. Which trade-offs you will accept, what you can tolerate being bad at for six months, which risk keeps you awake, what you would regret. None of that is inferable from a description of the subject, and a model asked to prioritise will produce something reasonable-sounding and generic, which is worse than nothing because it looks like an answer.
Minty Path reflects this in where it stops. AI proposals are reviewed change by change and approved individually; nothing is applied wholesale. Focus — the one-to-three branches getting attention — is never set by a proposal. And a branch cannot be marked Proven without real evidence attached, so a generated plan claiming parts are already done imports those rows as in-progress instead.
Those are deliberate boundaries: the machine drafts structure, you decide what matters and what is finished.
Getting the result into a map
Minty Path accepts id,parent_id,title,description,status,area,priority,evidence and previews before creating anything. The server validates unique IDs, that parents exist, absence of cycles, and depth and branch limits, so a malformed generation fails at preview rather than half-importing.
You can keep the model at arm's length — Minty Path builds the prompt, you run it wherever you like, you bring the CSV back, and nothing of yours is transmitted. Or signed-in users can connect their own OpenAI, Anthropic, or Gemini account, with credentials encrypted under a server keyring outside the web root that never returns to the browser, exports, diagnostics, or backups.
Try the workspace without an account first — it is worth knowing what a good map feels like before asking a model to produce one.