quill-deep-analysis

Quill Deep Analysis

A Claude Code skill that does deep map/reduce research over your Quill meeting corpus and writes a fact-first analytical report you can drill into.

Ask one long-span question — "how did our enterprise pillars hold up over the last six weeks?", "give me a balanced overview of Jerry's last year for a performance review" — and it shards the timeline across parallel sub-agents, consolidates the evidence per dimension, and produces a report with a dated timeline, finding-stating sections, and linked sub-reports down to the underlying meeting quotes.

It is not for single-meeting recaps or single-fact lookups — query your meetings directly for those.

How it works

MAP (by time window) → REDUCE (per dimension) → SYNTHESIZE (top-level report) → RECURSE (on generated questions)

• Map — one sub-agent per time window reads that window's transcripts and extracts cited evidence for every dimension at once.
• Reduce — one sub-agent per dimension consolidates across all windows and narrates how things evolved.
• Synthesize — one sub-agent writes the top-level report (00-SUMMARY.md) and generates follow-up questions.
• Recurse — the sharpest questions get a targeted second research pass.

The artifact mirrors the computation, so you can drill 00-SUMMARY → <dimension>/report → _raw/window-NN and trace every claim to a meeting link. A bundled report-viewer.html renders the tree with reader-view typography, light/dark, and click-through navigation.

Requirements

• Claude Code with the Workflow tool available.
• The Quill MCP connected (the skill uses mcp__quill__search_meetings, get_transcript, get_meeting, search_minutes). This skill is built for the Quill meeting corpus specifically.

Install

/plugin marketplace add quillmeetings/claude-plugins
/plugin install quill-deep-analysis@quillmeetings

(Installs from the Quill Meetings marketplace — one add, then any Quill skill.) Or install manually by copying skills/meeting-transcript-research/ into ~/.claude/skills/ and workflows/map-reduce-research.mjs into your project's .claude/workflows/.

Validate locally before publishing:

claude plugin validate .
claude --plugin-dir .

Permissions

The sub-agents write report files. Interactive runs are prompted normally, but background/headless sub-agents can't answer a prompt — so for a clean experience, allow Write/Edit scoped to your working tree:

// .claude/settings.json
{ "permissions": { "allow": ["Write(<project-root>/**)", "Edit(<project-root>/**)"] } }

Plugins can't grant permissions, so add this yourself. Without it the skill still works — agents return their report text and the caller persists it — just with an extra step.

Usage

Just ask a long-span, multi-meeting research question and the skill plans the decomposition. The only required input is the question.

Under the hood it runs workflows/map-reduce-research.mjs via the Workflow tool with:

arg	meaning
question	the top-level question (string)
dimensions	[{ key, name, definition, probes }] — put a scoring rubric in a definition if you want one; the engine won't invent a scale
windows	[{ start, end, label }] — pre-sharded ISO date-times
participantsHint	names usually present (context, not a hard filter)
outDir	run folder, e.g. research/<topic>-<date>
recursion	{ rounds: 1, topK: 4 }, or { rounds: 0 } to stop after synthesis
models (optional)	per-agent overrides; default { map: 'sonnet', reduce: 'opus', synthesize: 'opus', recurse: 'sonnet', finalize: 'sonnet' }

See skills/meeting-transcript-research/SKILL.md for the full methodology, Quill query mechanics, and the report-register rules.

License

Apache-2.0. See LICENSE.