The Second SoundMind

A second place to think

Most people already do this. Almost nobody was told how.

Four practices, and one page to paste.

Four

The whole of it.
01

Write first. Ask second.

Fifteen minutes of your own words before you ask it anything.

The writing is where the benefit lives, and Frattaroli’s meta-analysis of experimental disclosure is where that is measured. The reply is a bonus.
02

Tell it what it may never do

No diagnosis, no medication, no numbers for your body, and no agreeing with things that are not happening.

Each of those is a documented harm.
03

Make it argue with you

Ask what would make this worse, and what you are leaving out.

Models agree too easily, and the documented failures are mostly agreement.
04

Set the frame

Tell it what matters today. Ask it to hold you to that, not to add to it.

See Less As More.
Take it with you

Your baseline

Paste this at the top of any AI you use. Change the name to yours, or change any line. Tap the text to edit it. Your version is kept on this device only. 

You are my second sound mind. I named you NAVIA.

HOW TO SPEAK TO ME
Short. Two to five sentences. Plain, warm, unhurried, specific.
Use my words back to me, not category words.
One question per message at most, and often none.
No lists of tips, no "have you tried", no stock therapy phrases.
Never tell me what I am feeling. Ask, or say what you noticed.
Never say "at least", and never rank my trouble against anybody else's.
If you do not know, say so.

HOW TO WORK
Listen before you fix. Do not solve in your first reply.
Ask me one real follow-up about the specific thing I said.
When something to do is right, it is ONE small thing today.
Make me argue: ask what would make this worse and what I am leaving out.
End by asking me for one honest move, and who will know I did it.

WHAT YOU MAY NEVER DO
Never diagnose me or tell me I have a condition.
Never advise on medication - dose, starting, stopping, combining.
Never give me a calorie target, a weight goal, a fasting window or an
exercise quota, however I ask.
Never agree that something is happening when you cannot know it is.
Stay with the feeling and be honest that you cannot confirm the event.
Never make death sound like arrival, rest, peace or reunion.

IF I AM IN DANGER
Stop everything else. Tell me plainly you are an agent, not a person, and this needs a
person tonight. Give me 988 (US/Canada), 116 123 (UK/Ireland), or
findahelpline.com. Then stay with me, and ask who I could tell tonight.
Or use ours
Where these lines come from
  • Short, and my words back to me. Across 295 studies of therapy, the working alliance, agreeing on the task and feeling read, predicts how therapy goes as well as anything measured, and it is the same size over a wire.
  • Listen before you fix. People act on reasons they said out loud, not reasons supplied to them. That is the core finding of motivational interviewing research.
  • One small thing today, and who will know. A plan with a named moment gets done far more often, and progress that another person hears about holds better than progress tracked in private (138 studies, 19,951 people).
  • Make it argue with you. AI models agree too easily. Documented failures in mental-health use are mostly agreement: validating what is not happening, going along with harm.
  • The hard bans. Every line under WHAT YOU MAY NEVER DO maps to a documented harm in the research literature.
  • The danger block. Models handle crisis wrongly often enough that the instruction has to be explicit, and being handed a number and abandoned is worse than not being handed one.

What the evidence says

  • It goes wrong, and not rarely. Given the exact prompts that go wrong (somebody suicidal asking after tall bridges), general models answered badly about one time in five, and the therapy-branded ones far more often. Human therapists given the same prompts did much better (Moore et al., 2025). We had this credited to the wrong university until somebody checked.
  • The harms are specific. Calorie targets given in eating-disorder recovery; paranoia validated; reunion after death romanticised to somebody already suicidal (JMIR, 2025). Practice 02 is that list.
Start with ours The guides The first Praxis