Document longer persona variants and post-release dialog composition

This commit is contained in:
Jacob Dubin
2026-05-17 20:29:39 -05:00
parent af76cbaee2
commit 9ffdd6d09e
2 changed files with 39 additions and 0 deletions

View File

@@ -836,6 +836,7 @@ Current release theme:
- personality and capability questions - personality and capability questions
- favorite-style prompts like `what is your favorite color` - favorite-style prompts like `what is your favorite color`
- attraction and preference prompts like `what is your favorite flower`, `do you like R2D2`, `do you like the sun`, `do you like space`, and `do you like kids` - attraction and preference prompts like `what is your favorite flower`, `do you like R2D2`, `do you like the sun`, `do you like space`, and `do you like kids`
- longer authored variants for the same prompt family when Pegasus shows richer phrasing
- charm/capability prompts like `can you laugh` and `can you dance` - charm/capability prompts like `can you laugh` and `can you dance`
- mood / affect questions - mood / affect questions
- recognition follow-ups like `do you know me` - recognition follow-ups like `do you know me`
@@ -846,6 +847,7 @@ Current release theme:
- keep adding small sourced personality batches, especially the legacy `R2D2`, `sun`, `space`, `kids`, and charm prompts - keep adding small sourced personality batches, especially the legacy `R2D2`, `sun`, `space`, `kids`, and charm prompts
- keep adding 1-3 persona prompts per pass with tests - keep adding 1-3 persona prompts per pass with tests
- prefer source-backed MIM imports when the legacy text is available, and use a temporary runtime reply only when needed to unblock user value - prefer source-backed MIM imports when the legacy text is available, and use a temporary runtime reply only when needed to unblock user value
- keep a separate note for longer authored variants so we do not lose the multi-clause Peggy-style phrasing while importing the short-form packs
- Mood follow-up work in flight: - Mood follow-up work in flight:
- source-backed happy/sad/angry response packs are now part of Build B - source-backed happy/sad/angry response packs are now part of Build B
- small-talk aliases like `what are you up to` and `how are things` now stay on the emotion-query path - small-talk aliases like `what are you up to` and `how are things` now stay on the emotion-query path
@@ -868,6 +870,38 @@ Current release theme:
- each pass can be scoped to a small batch of prompts - each pass can be scoped to a small batch of prompts
- the backlog makes it obvious what is still missing without losing momentum - the backlog makes it obvious what is still missing without losing momentum
### 31. Longer Authored Persona Variants
- Status: `ready`
- Tags: `content`, `docs`, `protocol`
- Why now:
- Pegasus often used longer, multi-clause authored alternatives for the same personality question
- we already have the short-path import working, so this is a low-risk way to add richer phrasing without inventing a new dialog engine
- it gives us a straightforward next pass that stays familiar to the original robot
- Scope:
- import the longer authored variants already present in the legacy MIMs
- prefer richer phrasing for favorite-style, identity, and charm prompts when the source text provides it
- keep the runtime behavior rule-based and deterministic
- Next step:
- add a small batch of longer variants to the current Build B content packs and prove them with a smoke test
### 32. Dialog Joining And Composition
- Status: `discovery`
- Tags: `content`, `docs`, `protocol`
- Why now:
- the videos and source files suggest Jibo sometimes felt like he was joining thoughts together, even when the source text was still authored
- we have not found evidence of a general runtime joiner yet, so this remains a post-release enhancement instead of a 1.0.19 dependency
- keeping it separate lets us preserve familiar Jibo phrasing now and experiment with composition later
- Scope:
- design a post-release dialog composition layer that can stitch authored fragments together when appropriate
- keep the first version conservative and familiar, not LLM-driven
- make sure any future joining feature is opt-in and does not replace the current authored prompt path
- Follow-up:
- revisit after 1.0.19 personality import and report-skill parity stabilize
- decide whether the composition layer should sit above the prompt catalog or beside it as a dedicated response post-processor
- keep this separate from the authored-variant backlog item so we do not blur prompt richness with runtime composition
## Suggested Order ## Suggested Order
Before closing `1.0.18`: Before closing `1.0.18`:
@@ -903,6 +937,8 @@ For `1.0.19`:
15. Grocery list capability discovery and MVP selection 15. Grocery list capability discovery and MVP selection
16. Lasso, identity, and onboarding as larger discovery-driven tracks 16. Lasso, identity, and onboarding as larger discovery-driven tracks
17. Legacy MIM personality import ladder and first declarative prompt packs 17. Legacy MIM personality import ladder and first declarative prompt packs
18. Longer authored persona variants for the same prompt families
19. Dialog joining/composition as a post-release enhancement, kept separate from the 1.0.19 ladder
For `1.0.20` and beyond: For `1.0.20` and beyond:

View File

@@ -29,6 +29,7 @@ Keep a running checklist of the legacy persona questions and identity surfaces w
- self-description and social charm: `what's your name`, `do you have a nickname`, `do you like being Jibo`, `are there others like you` - self-description and social charm: `what's your name`, `do you have a nickname`, `do you like being Jibo`, `are there others like you`
- favorite-style prompts: `what is your favorite color`, `what is your favorite food`, `what is your favorite music` - favorite-style prompts: `what is your favorite color`, `what is your favorite food`, `what is your favorite music`
- attraction and preference prompts: `what is your favorite flower`, `do you like R2D2`, `do you like the sun`, `do you like space`, `do you like kids` - attraction and preference prompts: `what is your favorite flower`, `do you like R2D2`, `do you like the sun`, `do you like space`, `do you like kids`
- longer authored variants for the same prompt family when Pegasus shows richer phrasing, especially multi-clause and follow-up-heavy responses
- capability and charm prompts: `can you laugh`, `can you dance` - capability and charm prompts: `can you laugh`, `can you dance`
- affect and mood: `how are you`, `are you happy`, `are you sad`, `are you angry` - affect and mood: `how are you`, `are you happy`, `are you sad`, `are you angry`
- memory and identity recall: `who am i`, `what is my name`, `when is my birthday`, `what is my favorite music` - memory and identity recall: `who am i`, `what is my name`, `when is my birthday`, `what is my favorite music`
@@ -228,6 +229,8 @@ Port after Build B:
- richer `scripted-responses` families that depend on follow-up state - richer `scripted-responses` families that depend on follow-up state
- special-date / holiday personality sets - special-date / holiday personality sets
- more nuanced chitchat branches that need context-aware routing - more nuanced chitchat branches that need context-aware routing
- longer authored variants for existing prompts when the source text contains them, so the robot keeps the familiar Pegasus cadence without inventing new dialog composition yet
- dialog joining / composition as a post-release feature, kept out of the 1.0.19 ladder so we do not blur authored phrasing with a runtime joiner
Why these later: Why these later: