Two layers
Vibely separates “who you are as a builder” from “who the agent is to you.” Both compose; both ship to the model on every turn.| Layer | What it is | Max length |
|---|---|---|
| Custom instructions | Facts and preferences about you and how you like the agent to work. | 4000 chars |
| Customize Vibely | The agent’s persona, intro line, personality tags, extra info. | ~2000 chars |
Custom instructions — write them like a colleague handover
The agent reads this on every turn, in every project, forever. Treat it like the handover note you’d give a new contractor on day one:- What you’re building (one line)
- The stack you prefer when it’s not specified
- Anti-preferences (don’t suggest X, don’t use Y)
- How verbose you want explanations
- Your name and what to call you
Example that works
Example that doesn’t
What to put in (and what not to)
Put in:- Stack preferences when “default to X” actually changes behaviour
- Anti-patterns you’ve been bitten by (“don’t use useEffect to derive state”)
- Conventions specific to your projects (“commit messages in conventional commits format”)
- Personal context only relevant to how the agent talks (“I’m dyslexic, prefer code blocks over prose”)
- Generic best-practice (“write clean code”) — the model already tries to
- Per-project file paths or function names — those belong in project memory, not your account
- Secrets or anything you don’t want shipped to every model provider on every turn
Customize Vibely — the agent’s persona
This is the smaller sister panel. It changes how the agent introduces itself and speaks, not what it knows about you.- Intro line — what the agent opens new chats with. Defaults to a friendly Vibely line; replace with something terser if you want.
- Personality tags — pick from a chip list (e.g. “concise”, “playful”, “technical”). The agent tries to match the picked vibe.
- Additional info — a small free-form field the agent reads as system context. Use it for anything that doesn’t fit the tags.
- Enable for new chats — when off, the persona only applies to chats where you explicitly turn it on. Default is on.
How it reaches the model
Both layers are merged into a memory block that prepends every turn — same priority as the system prompt. The agent sees them before it sees your message. This means:- Conflicts resolve in your favour. If you said “don’t use Tailwind” and the prompt asks “build a Tailwind page”, the agent will surface the conflict instead of silently picking one.
- The block is not model-visible across providers — it lives on Vibely’s backend and is injected per turn.
- Editing instructions takes effect on the next turn, not retroactively to the conversation so far.
