Skip to content

scratchpad.md

Latest commit

 

History

History
176 lines (132 loc) · 7.6 KB

scratchpad.md

File metadata and controls

176 lines (132 loc) · 7.6 KB

Scratchpad

Quick capture for build ideas, questions, and TODOs. Review periodically; move mature items to dev/work/plans/ or turn into PRDs.

Ideas & Raw Captures

[2026-03-05] profile.md as silent hard dependency for CommitmentsService

Discovered during smoke test: if context/profile.md doesn't exist, ownerSlug is undefined in entity.ts, and parseActionItemsFromMeeting is silently skipped — meaning arete commitments list always returns empty with no warning. The entire CommitmentsService producer path is gated on this file. Fresh workspaces that skip profile setup will have broken commitments silently.

Fix candidates:

  • Prompt for profile during arete install / arete setup (name, role, email)
  • Add a guard in arete people memory refresh that warns when profile.md is missing
  • Add arete status check for missing profile.md

Small plan. Add to dev/work/plans/ when ready to scope.

[2026-02-12] User voice / writing style from uploads

User said: "I'd like the user to be able to upload emails, PRDs, documents, etc. to help define and shape their voice for the agent." Explore: let users upload sample artifacts so the agent can learn and mirror their tone, structure, and style. Could be a "voice" context folder, a briefing step that injects style cues, or an onboarding flow. Needs product exploration.

[2026-02-12] Meeting agendas shouldn't be so robust

Simplify prepare-meeting-agenda output; avoid over-engineered or overly detailed agendas.

Skills discovery on no-match

When the router returns no match, Areté could run npx skills find <query> and suggest installable skills from skills.sh. Explore: integrate into the "no match" path or add a "suggest skills" step in GUIDE mode.

[2026-03-19] Brief feature needs better summaries

Currently arete brief --for "topic" returns relevance-ranked file pointers, but summaries are often just filenames or titles (e.g., "Project doc: account-rollout-analysis.md"). The agent still has to read files to understand them.

Issues identified:

  • extractSummary() pulls first paragraph, but many files have minimal/unhelpful openings
  • Project docs often show as just "Project doc: filename.md"
  • Meetings show date/source metadata instead of what was discussed
  • People files show role/company but not relationship context

Improvement candidates:

  1. Smarter extraction — Pull first substantive paragraph (skip metadata, headers, HTML comments) + key bullet points
  2. Structured frontmatter — Add optional summary: field that gets prioritized
  3. Pre-computed summaries — Generate AI summaries at index time (cost vs. quality tradeoff)
  4. Entity-aware formatting — Format people/meetings/projects differently with relevant fields
  5. Configurable depth — --brief (pointers only) vs --detailed (include content snippets)

The value of good summaries: agent can answer simple questions without file reads, faster context assembly.

[2026-03-19] Meeting Intelligence Improvements

Inspired by reviewing skills.sh/meeting-minutes. That skill has a strict 12-section schema with acceptance criteria on action items, parking lot, risks/blockers, etc. We have better intelligence (memory integration, people entities, commitments) but their output quality for shareable notes is stronger.

Three distinct improvements identified:

1. Better Overview/Summary

Current summary is 2-4 sentences, basic. Could be more structured:

## Summary
**Outcome**: Aligned on Q2 API priorities; unblocked JWT auth decision.
**Key Topics**: API versioning (decided), JWT auth (decided), Pricing model (parked)
**Open Questions**: How does this affect mobile SDK timeline?

Scope: process-meetings extraction enhancement.

2. Shareable Export (arete meeting share)

Pain point: processed meetings aren't easily shareable (Slack, email).

Proposal: New CLI command + UI action (post-processing, not automatic)

  • CLI: arete meeting share <file> or arete meeting export <file> --format slack|email|markdown
  • UI: "Share" button in arete view on processed meetings

Output format — clean, portable:

# Glance Email Sync — Mar 19, 2026
## Summary
Aligned on Q2 API priorities. Decided JWT for v2 auth.
## Decisions
| Decision | Approved By | Effective |
| Use JWT for v2 | Sarah, Mike | Q2 launch |
## Action Items
| Action | Owner | Due | Done When |
| Draft JWT RFC | @mike | Mar 22 | RFC reviewed by security |
## Parking Lot
- Pricing model → needs finance input
---
*Generated from [Meeting](link) via Areté*

3. Recurring Meeting Continuity

Connect meetings in a series so meeting prep can pull parking lot, open items, etc.

A. Meeting → Goal Linkage (optional frontmatter)

goal: q2-api-launch  # links to goals/q2-api-launch.md
  • Set manually, inferred during processing, or via meeting-prep when creating agenda from goal
  • Goals = stable anchor (Q1 roadmap), Projects = bounded work packages
  • Default to goals for recurring syncs

B. Parking Lot Extraction (new staged section type)

## Staged Parking Lot
- pk_001: Pricing model (needs finance input)
- pk_002: Mobile SDK timeline (discuss next sync)

On approval: if meeting has goal: → append to goal's open-questions. Else → stays in meeting file for next in series.

C. Meeting Series (inferred, not explicit) Don't create "series" entity — use title similarity:

  • arete search "Glance Email Sync" --scope meetings --days 30
  • Pull parking lot + open action items from most recent match
  • Include goal context if linked

Meeting prep flow with continuity:

User: "Prepare an agenda for Glance Email Sync"
Agent:
1. Search recent meetings with similar title
2. Extract parking lot + open action items
3. Include goal context if linked
4. Generate agenda with carryover items

Implementation order (suggested):

  1. Parking lot extraction (unlocks #3) — Medium
  2. Meeting → Goal linkage — Low
  3. arete meeting share CLI + UI — Medium (immediate pain point)
  4. Meeting prep pulls prior meetings — Medium
  5. Better summary extraction — Low (polish)

These are related but separate plans. Start with share command (user pain point) or parking lot (architectural foundation).

[2026-03-19] Commitments Review Skill/UI

Review and organize commitments with agent assistance.

Two-part experience:

A. Agent-Assisted Cleanup Agent reviews open commitments and recommends a few to drop:

  • Likely completed (context suggests resolved)
  • Stale (no activity, old creation date)
  • Superseded (later commitment covers same thing)

Agent surfaces recommendations with reasoning; user approves/rejects each.

B. Quick Priority Organize Let user quickly organize remaining commitments:

  • Priority tiers (this week / soon / someday)
  • Bulk actions (mark done, snooze, drop)
  • Filter by person, project, or age

Could be CLI (arete commitments review) or UI tab in arete view.

[2026-03-19] Meeting Review Skill

Pick a few meetings from the past week and have an agent review + provide feedback.

Use cases:

  • Self-coaching: Did I run good meetings? What could improve?
  • Pattern recognition: What themes/decisions are repeating?
  • Meeting hygiene: Are outcomes clear? Action items actionable?

Possible outputs:

  • Per-meeting feedback (structure, outcomes, what was effective/missing)
  • Cross-meeting synthesis (themes, recurring topics, relationship health signals)
  • Suggestions for follow-ups or prep improvements

Could tie into week-review skill or be standalone.