I think all the prompting (system prompt, presets, skills, ...) will be rapidly evolving.
So how about we decouple all the prompting from the rest of the code? Maybe as standalone files (prompts/**/*.md, prompts/sytem_prompt.md, prompts/presets/*.md, prompts/skills/*.md, ...).
The idea here is that we can rapidly adapt the prompting, while keeping the rest of the case (almost) intact.
For example, breaking down prompt injection in pre-defined phases => prompts hook into these phases => decoupling.
Not a blocker, just an optional optimization.
(In the long-term, we could even enable users to customize all the prompting. Clearly not a priority for now, and clearly post-MVP. Potential vision: build a flexible architecture for all kinds of autopilot AI use cases — abstract idea, not sure if it has real use cases.)
I think all the prompting (system prompt, presets, skills, ...) will be rapidly evolving.
So how about we decouple all the prompting from the rest of the code? Maybe as standalone files (
prompts/**/*.md,prompts/sytem_prompt.md,prompts/presets/*.md,prompts/skills/*.md, ...).The idea here is that we can rapidly adapt the prompting, while keeping the rest of the case (almost) intact.
For example, breaking down prompt injection in pre-defined phases => prompts hook into these phases => decoupling.
Not a blocker, just an optional optimization.
(In the long-term, we could even enable users to customize all the prompting. Clearly not a priority for now, and clearly post-MVP. Potential vision: build a flexible architecture for all kinds of autopilot AI use cases — abstract idea, not sure if it has real use cases.)