Reported via beta feedback (v0.5.1).
A user authoring markdown documents with the agent wants a way to give persistent directives about how the model formats its output -- and there's no obvious place to set one.
Concrete motivating case: the model emits GitHub-flavored markdown that omits blank lines around lists. The user's external markdown editor follows stricter CommonMark rules and won't recognize a list unless it's surrounded by blank lines, so lists render as plain paragraphs. They'd like to say once -- "write CommonMark-compatible markdown, blank line before and after every list" -- and have it stick for the rest of the session/project instead of re-typing it each prompt.
The ask, generalized: a persistent user-directive / "custom instructions" mechanism that gets injected into the agent's context -- a place to record house-style rules (list spacing, tone, formatting conventions) so they aren't repeated every turn and don't get dropped as the conversation grows. Today the system prompt is entirely code-defined (extensions/loom/context.ts); there's no user-configurable surface that feeds into it.
Why it matters: users are producing real deliverables (documents, notebooks, posts) with Orbit, and small formatting mismatches like this quietly break downstream tooling (editors, renderers). A one-time directive surface is low-friction and generalizes well past this single case.
Possible directions:
- A project-level instructions file (an
AGENTS.md-style doc always loaded into context), if one doesn't already exist.
- A "Custom instructions" field in Preferences that's appended to the system prompt.
- As a cheap stopgap independent of the above: nudge the default system prompt to prefer CommonMark-safe markdown (blank lines around lists), which renders correctly under both GFM and strict parsers -- so the common case is right without any user action.
(Beta tester feedback; details anonymized.)
Reported via beta feedback (v0.5.1).
A user authoring markdown documents with the agent wants a way to give persistent directives about how the model formats its output -- and there's no obvious place to set one.
Concrete motivating case: the model emits GitHub-flavored markdown that omits blank lines around lists. The user's external markdown editor follows stricter CommonMark rules and won't recognize a list unless it's surrounded by blank lines, so lists render as plain paragraphs. They'd like to say once -- "write CommonMark-compatible markdown, blank line before and after every list" -- and have it stick for the rest of the session/project instead of re-typing it each prompt.
The ask, generalized: a persistent user-directive / "custom instructions" mechanism that gets injected into the agent's context -- a place to record house-style rules (list spacing, tone, formatting conventions) so they aren't repeated every turn and don't get dropped as the conversation grows. Today the system prompt is entirely code-defined (
extensions/loom/context.ts); there's no user-configurable surface that feeds into it.Why it matters: users are producing real deliverables (documents, notebooks, posts) with Orbit, and small formatting mismatches like this quietly break downstream tooling (editors, renderers). A one-time directive surface is low-friction and generalizes well past this single case.
Possible directions:
AGENTS.md-style doc always loaded into context), if one doesn't already exist.(Beta tester feedback; details anonymized.)