STOA

Anxiety disrupts the ability to think clearly and act, yet existing tools wait for the user to open an app and start an exercise at exactly the moment they are least able to. Most people never do; avoidance is the most common response.

STOA reads the body’s stress signals and starts an intervention on its own, cognitive, behavioral, or body-based, matched to the moment, closing the gap between needing help and being able to seek it.

A panic episode is a behavior pattern, not a demographic, so the empathy map records what someone says, thinks, does and feels during an episode rather than who they are. The pains and gains are written as patterns, loss of agency, privacy and dignity, return to baseline, so they hold for anyone having this experience.

Two jobs came out of it. The primary one is what the product is built to do. The secondary one turned out to matter almost as much: people cannot always tell a panic attack from a heart attack, and the fear of getting that wrong is its own source of panic.

Working solo with an agent, the bottleneck is not producing screens, it is keeping fifty of them consistent. So the visual and interaction rules went into a single CLAUDE.md brief before any screen was generated, and every screen was checked against it afterward.

Where AI did the producing, the direction stayed mine: Elicit synthesized the literature and I verified every finding before using it; an image model drew the storyboard panels from descriptions I wrote; Claude Code built screens against the brief. The palette, the type, the layout and the interaction patterns came from the research.

The detection half of this product cannot be tested on a screen, so it was tested on a wrist. I built a band with an LED and drove it by hand while someone wore it: amber when I decided their stress had spiked, green once the exercise brought them back. A Wizard-of-Oz rig, where the intelligence is a person hiding behind the curtain.

Then I put the work in front of a critic and asked an agent to attack it. Five things came back, and each one changed something:

1. SourceThe critiqueWhat changed
2. Design crit

   The problem statement was too solution-specific, locking the work into Stoic interventions before any divergent exploration could happen.

   Rewrote it as the challenge that now closes Research, dropping three assumptions: that every intervention is Stoic, that cognitive reframing is always the mechanism, and that users already have coping strategies.

3. Design crit

   The empathy map read like a persona. Pains and gains named exact time windows and treated colleagues as the only audience for disclosure.

   Rewrote both as behavioral patterns, managing episodes with privacy and dignity, regaining control more quickly, so they describe the episode rather than one person.

4. Design crit

   Stoicism was carrying more weight than the evidence supports.

   CBT became the framework and Stoicism became the voice. The techniques are the clinically grounded ones; the tone is what makes them land.

5. Agent critique

   “I feel calmer” implies a binary success state, with no path for someone who feels worse.

   The check-in now offers five honest options including “still anxious.” Where that path leads is still unbuilt, and it is the first thing in the backlog.

6. Agent critique

   Onboarding shows a live heart-rate dashboard without ever asking to read the wearer’s body.

   Named as a gap rather than patched. A product that reads your body has to ask first, and the consent flow is the second thing in the backlog.

STOA notices before the wearer does, so the first thing anyone sees is on their wrist rather than in their hand. Three states carry it: a resting baseline, the reading that crosses the line, and the anchor offered straight away, all reachable without taking a phone out in a room full of people.

1. 01Darken Pewter until it clears AA

   Pewter is the token for body and secondary text, and on Marble White it measures 4.41:1, just under the 4.5:1 that AA requires. On pure white it reaches 4.62:1 and passes, which is how it survived review. Darkening it to #57757D puts it at 4.72:1 on both grounds, a shift small enough that nothing else in the system has to move.

2. 02Stop treating Calm Blue as a text color

   Calm Blue measures 3.52:1 on Marble White. That clears AA for large text only, so it works as an accent, an icon fill, and a chart color, and should never be asked to carry a sentence. The style guide currently lists it next to Pewter as body and secondary.

3. 03Keep one card radius, not two

   Cards are specified at 24px in one place and 16px in another, under names close enough to be mistaken for each other. Pick one radius for cards and give the sheet its own name, and the rule becomes something another designer can follow without asking which card this is.

4. 04Retire the superseded type sheet

   The Figma file still holds an earlier typography sheet set in Georgia alongside the current one set in Cormorant Garamond. Both are reachable, neither is marked, so the first thing a new person opens is a coin flip. Delete the old one.