| name | what-do-i-think |
| description | Synthesize the user's views on a topic across 10+ conversations. Modes synthesize (default) or precedent. Use when user is forming an opinion and may have already concluded something. |
| layer | L2 |
| reads | ["L2.summaries","L1.embeddings"] |
| writes | [] |
| citations | required |
| determinism | temporal |
| allowed-tools | mcp__my-brain__what_do_i_think |
what-do-i-think — L2 opinion synthesis
Framework context
L2 synthesis that surfaces what the user has already decided/concluded about a topic before they re-ask. Deduplicates by 80-char lowercase prefix. Orders by importance (breakthrough > significant > routine). Every decision/question/quote carries a citation.
Precedent mode is a narrower projection: returns the closest past matches for a situation, filtered to exclude "none identified" decisions.
When to invoke
- User says "what do I think about X" / "have I decided on Y"
- User is about to make a decision — surface their own prior thinking first
- Mode
precedent when user cites a specific situation ("I'm facing a decision like Z")
Input
topic: str
mode: "synthesize" | "precedent" = "synthesize"
Output contract
Synthesize mode:
## What you think about: {topic}
### Top 5 summaries
{icon} **{title}** [conv_id · date]
> {summary_300ch}
### Key decisions (dedup by 80ch prefix, top 10)
- {decision} [conv_id · date]
### Still open (top 8)
- {question} [conv_id · date]
### Authentic quotes (top 5)
> "{quote}" [conv_id · date]
Precedent mode: top 10 closest matches, each row cites one summary with decision field.
Does NOT do
- Declare "you think X" as a synthesized claim without a citation trail
- Merge contradictory views silently — surface both
- Paraphrase the decisions (verbatim, truncated to 200ch)
Fallback
No summaries → Lance + keyword search on L0. Appends: "Running without summaries. For richer synthesis: brain-mcp summarize"
Verification checklist
