Some weeks I ship everything on my list. Other weeks I stare at a half-written email for two hours and call it a day. I used to blame focus, sleep, or motivation — until I started treating my decisions like data instead of feelings.
That habit turned into a product: an AI-assisted decision report engine that frames a structured analysis as a tarot-style reading. This post is about the model underneath — no mysticism required — and what building it taught me.
The observation
For years, whenever I faced a decision I was avoiding, I'd pull tarot cards. Not for prediction — but because a spread forces you to look at the situation from angles you'd rather skip. Positions like "what you fear" or "what you refuse to see" are basically canned journal prompts.
I noticed I kept asking the same seven questions. So I wrote them down, scored them, and mapped them onto continuous dimensions.
The seven dimensions
Every decision report reduces the situation to seven indices on a 0–100 scale (in practice calibrated to an observed 12–88 band, because humans rarely live at the extremes):
| Dim | Name | What it measures |
|---|---|---|
| R | Risk appetite | tolerance for irreversible downside |
| I | Intuitive synthesis | how much you trust pattern-recognition vs. explicit reasoning |
| P | Temporal patience | willingness to wait for compounding |
| S | Social dynamic | how much alignment/consensus you need |
| A | Action bias | commit-fast vs. deliberate |
| Rf | Reflection depth | tendency to re-analyze |
| St | Stability | preference for known structures over flux |
A quiz (about 3 minutes) scores the question, not the person. The output is a seven-dimension radar plus the two or three tensions hiding inside it — e.g. high reflection + low action under load reads as "decision fatigue wearing a costume."
Archetypes as geometry
Ten recurring vectors fell out of the calibration data — each is just a shape across the seven axes:
| Archetype | R | I | P | S | A | Rf | St | Failure mode |
|---|---|---|---|---|---|---|---|---|
| Alchemist | 66 | 22 | 58 | 30 | 64 | 44 | 34 | solitary conviction without alignment |
| Sage | 40 | 78 | 74 | 52 | 20 | 72 | 56 | over-analysis; missed asymmetric upside |
| Challenger | 76 | 40 | 28 | 44 | 84 | 30 | 32 | hyperactive interference; burn rate |
| Builder | 44 | 62 | 74 | 52 | 50 | 58 | 76 | rigidity past optimal pivot points |
| Creator | 60 | 30 | 30 | 42 | 76 | 30 | 28 | draft overload; consolidation lag |
| Opportunist | 74 | 50 | 34 | 46 | 72 | 34 | 30 | shallow commitment; option chasing |
| Architect | 38 | 78 | 76 | 50 | 58 | 74 | 76 | inelasticity in fast-changing terrain |
| Leader | 56 | 56 | 64 | 82 | 72 | 52 | 64 | high consensus load; diluted velocity |
| Maverick | 76 | 40 | 30 | 22 | 78 | 30 | 26 | key-person isolation |
| Competitor | 70 | 50 | 40 | 58 | 78 | 40 | 48 | scoreboard fixation |
These aren't personality verdicts — they're operating modes. The same person flips vectors depending on the domain (work vs. relationships vs. money).
Engineering notes
- Quiz-gated queue. Free beta is 3 readings/day, gated by the quiz so the queue stays human-scale and bots can't burn the LLM budget.
- The LLM does framing, not fortune-telling. It renders tensions and next steps from the scored dimensions; the tarot layer is a presentation skin over a structured report.
- Honest statistics. The birth-timestamp mapping that seeds the archetype guess is a weak statistical correlation. The product says so, in the UI and in the copy. Users trust it more, not less, for that.
- Naming is the intervention. The most common feedback: "I froze less once the pattern had a name." A labeled failure mode is fixable; unnamed anxiety isn't.
What I'd do differently
- Ship the radar before the tarot skin. The mystical frame is great for reach, but the retention comes from the structured report.
- Score the question first, person second. Question-scoring generalizes; profiling feels invasive.
- Ship with the disclosure. Every A/B test said honesty converts better than mystique.
If you're stuck in a decision loop, try tracking one dimension — risk or action — for a week. Naming the pattern is half the fix.
I'm building this in public — the engine is in free beta (tangoera.com) and the longer build log lives on Indie Hackers. Happy to answer questions about the calibration data or the quiz design.
Top comments (0)