The tool named itself. Synantics is a compound of synthesised semantics, and it came out of the system's own six stages — which is the shortest available proof that the process works.
A name is the output of a decision sequence, not of a prompt. Synantics makes that sequence explicit: six stages, each one approved by the user before the next begins, so that by the time names are generated the meaning, the boundaries and the evaluation criteria already exist.
Synantics is live. For the best results, add your own answers along the way. The more personal the input, the better the outcome.
Note: Please allow a couple of seconds for Render to load.
The problem
It began as Brandon, a custom GPT for brand naming, and I found the experience neither creative nor enjoyable. The questions blurred into one long prompt, there was no sense of progression and it was impossible to see how an insight became a territory and then a name.
Commercial AI namers have the same shape: type keywords, receive a list. The brief — audience, category, tone, direction, avoid-lists — is what actually determines name quality, and nobody enjoys filling in a twelve-field form to produce one.
How the system works
Six connected stages. Each combines fixed structure with custom input, and at every step the system proposes an interpretation that the user can revise, reject or approve before moving on.
| Stage | What happens |
|---|---|
| Frame | Defines what is being named, who needs it, the before-to-after change and the plausible-but-wrong assumptions to exclude. Produces a strategic foundation — no names, no metaphors. |
| Distill | Reduces the material to what the name must carry: what is felt immediately, what is discovered later, what tension shapes it, what must never lead. Produces a strategic core and an avoid list. |
| Disrupt | Deliberately pushes past obvious category language — movement, behaviour, environments, unexpected associations and ideas that are attractive but strategically wrong. |
| Connect | Links approved fragments into recurring patterns and distinct constellations. Still not names — a structured strategic field. |
| Crystallize | Turns patterns into a naming brief: leading territory, supporting territory, strategic role, emotional quality, vocabulary families, boundaries. |
| Express | Generates candidates with rationale and sound logic, then reviews and scores them against the approved criteria. |
The rule underneath all six is that the raw input is never overwritten. AI normalisation is stored alongside the user's own words, and later stages consume only values the user has approved — never an unapproved summary.
Architecture
A full-stack web application. A vanilla JavaScript frontend carries the interactive experience — answers appear as selectable bubbles that assemble into clusters. A Node and Express backend holds sessions, territories, candidates, shortlists and generation batches in PostgreSQL, with Zod validating the structured output the model returns. The LLM key stays server-side; the browser never talks to the provider directly. Tests mock the model client entirely, so the suite runs without network access.
The final score is calculated deterministically from the approved criteria. The model proposes; the arithmetic is not the model's.
Tested against three deliberately different briefs
- Repair Club — a neighbourhood repair club and tool library, driven entirely by custom input, testing whether everyday language can become a distinctive naming direction. Taboo words included repair, fix, tool, eco, community, hub.
- Architecture AI Assistant — a workflow companion for architecture studios, which had to feel credible and architectural without sounding like software. Taboo words included AI, automation, agent, workflow, BIM, smart, architect.
- Botanical Evening Aperitif — a non-alcoholic aperitif that had to lead with desire and atmosphere rather than absence. Taboo words included sober, zero, clean, detox, wellness, alcohol-free.
Each case sets a naming character, an approach, a must-avoid list and a single success test — the question the result has to answer.
The six stages
My contribution
Sole author. I designed the six-stage model, wrote the stage specifications and the system prompt, built the Express backend and the PostgreSQL schema and implemented the frontend interaction.
What Synantics is not
- Not a name generator. Generation is the last stage, not the product.
- Not a legal or domain clearance service. Availability and conflict notes are directional only, and the interface says so — formal trademark and domain review remains a separate professional step.
- Not autonomous. No stage advances on the model's own judgement; every one passes through an explicit user approval.
- Not account-based. The live deployment runs sessions without sign-in or saved user accounts.
The interesting problem was not generation quality — it was traceability. Once each stage stored the user's raw answer, the model's interpretation and the approved value as three separate things, it became possible to look at a final name and walk backwards to the decision that produced it. That audit trail is the same requirement I keep meeting in architectural work, where a result nobody can justify is a result nobody can defend.