Understanding Assistant Families

Every StretchGPT conversation runs in an assistant family — a purpose-built flavor of the assistant tuned to a kind of work. Choosing the right family up front gives you sharper, more relevant answers. This article explains the nine families, how to pick one, and how families interact with the rest of StretchGPT.

The nine families

StretchGPT ships with nine families. The default is Core; if you don't choose one (or choose something unrecognized), your conversation falls back to Core.

Family	Best for
core	General questions, writing, summarizing, and everyday reasoning — the capable all-rounder.
code	Writing, explaining, reviewing, and debugging code and technical configuration.
research	Digging into a topic, comparing options, and structuring findings.
ops	Operational summaries, status briefs, runbooks, and process questions.
learn	Explanations, tutoring, and step-by-step learning at your level.
support	Customer-support triage, drafting replies, and classifying requests.
media	Working with media-related content and descriptions.
creative	Brainstorming, copywriting, naming, and creative drafts.
vision	Vision-oriented tasks and image-related reasoning.

Each family shares StretchGPT's core guardrails — accuracy first, direct answers, no invented facts — but leans into the tone and structure that suits its work.

Choosing a family

Click New chat.
Before typing your first message, select the family from the family picker.
Send your first message. The conversation is now pinned to that family for its lifetime.
Need a different family? Start a new conversation and pick it there — a single conversation stays in the family it was created with.

A realistic example

You're a support agent triaging a tricky ticket. Instead of the default Core assistant, you start a conversation in the support family and paste the customer's message:

"Customer says their export keeps failing at 80%. Classify the priority, suggest a first reply, and tell me the internal next step."

The Support family responds with a priority, a customer-ready reply, and an internal next step — the shape of answer that family is built for. If you'd asked the same thing in code, you'd get a more technical, debugging-oriented take.

How families interact with grounding and training

Grounding is family-aware in spirit: whatever family you use, StretchGPT still grounds answers in the StretchSearch index when evidence exists and cites sources as [n].
Feedback is tagged by family: when you rate an answer, the resulting training candidate is recorded against that conversation's family. This means the training pipeline can improve, say, the code family from code-conversation feedback without muddying the support family. See Rating Answers and the Feedback Loop.
Titles and history are per conversation, not per family: switching families means a new conversation, which gets its own auto-generated title and its own memory.

Tips

Match the family to the deliverable, not just the topic. A "write me marketing copy about our API" task belongs in creative, even though it's about code.
Core is a safe default. If you're unsure, start in Core — it handles almost anything competently.
Don't fight the family. If answers feel off-target, you may be in the wrong family; start a fresh conversation in a better-matched one rather than wrestling the current thread.

FAQ

Can I change a conversation's family after it starts? No — a conversation keeps the family it was created with. Start a new conversation to switch.

What happens if I pick a family that isn't recognized? StretchGPT safely falls back to Core.

Do all families cost the same or use the same model? All families run through the same governed provider chain (your proprietary model first, then a teacher fallback). The family shapes the assistant's behavior, not which providers are connected.