THOUGHTFUL · FIELD NOTES

How we build
Thoughtful.

Working notes from the team behind the product — on safety architecture, measurement design, and the choices that shape AI-assisted mental healthcare. Less formal than our Reports. More direct from the build.

Read the field notes

Fourteen Essays · 2026

The collection so far.

SAFETY

AI safety is more than crisis detection

On the clinical risk that only becomes visible when you read the whole conversation.

5 min readMelanie Nimmo

SAFETY

Where is the human in "human in the loop"?

On the difference between human oversight of an AI system and human clinical judgment in a person's care pathway.

5 min readMelanie Nimmo

CRAFT

The handoff is part of the product

On the engineering decisions behind an AI that can point beyond itself without inventing the next step.

5 min readNicky Li

PRINCIPLES

What should we expect from AI built for mental health?

On why the useful comparison is not "AI versus AI," but the system built around the conversation.

5 min readAlexa Varsavsky

CRAFT

"Maygentic"

On the unglamorous work behind faster shipping, higher quality, and a method that outlasted the team that built it.

5 min readChase Hall, Product @ Thoughtful

PRINCIPLES

What an AI should never decide on its own

On the single line in our internal documentation that does the most work.

5 min readMelanie Nimmo

MEASUREMENT

The five things we measure after every conversation

On building session-level evidence in a category that's used to waiting months for survey data.

5 min readElizabeth Shaffer

MEASUREMENT

What we do when our AI gets it wrong

On the override that keeps our resonance metric from lying to us.

5 min readNicky Li

SAFETY

Three layers of detection, because risk doesn't show up the same way twice

On the safety architecture that makes the rest of the product possible.

5 min readMelanie Nimmo

MEASUREMENT

Leading indicators vs. lagging indicators in mental health care

Why the industry's favourite metrics tell you about a problem six weeks after it started — and what to measure instead.

5 min readEmily Dworkin

SAFETY

Why we don't track keywords for crisis

On the difference between a system that's searching for words and a system that's understanding people.

5 min readElizabeth Shaffer

CRAFT

The setup screen we rebuilt because users told us to

On the two-week feedback loop that reshaped how new users meet our product.

5 min readRich Collins

DESIGN

Why 10-minute sessions outperformed longer ones

On what happened when we stopped imposing a session length on users and let them choose.

5 min readChase Hall

DESIGN

What makes a memory worth keeping

On the design decisions hiding behind a feature that sounds simple.

5 min readJuan Estela

How we buildThoughtful.

Fourteen Essays · 2026

AI safety is more than crisis detection

Where is the human in "human in the loop"?

The handoff is part of the product

What should we expect from AI built for mental health?

"Maygentic"

What an AI should never decide on its own

The five things we measure after every conversation

What we do when our AI gets it wrong

Three layers of detection, because risk doesn't show up the same way twice

Leading indicators vs. lagging indicators in mental health care

Why we don't track keywords for crisis

The setup screen we rebuilt because users told us to

Why 10-minute sessions outperformed longer ones

What makes a memory worth keeping

Get the next field notein your inbox.

How we build
Thoughtful.

Get the next field note
in your inbox.