Meeting Minutes and Brief Automation

Technical Pain Points in Meeting Minutes Management

In consulting projects, processing meeting minutes presents specific challenges: raw transcripts contain frequent homophone errors, particularly for management consulting jargon (e.g., "经营分析会" misheard as "精英分析会"); manual proofreading is time-consuming and prone to omissions; document formatting must adhere to strict standards (like the 附件10 template), but manual layout is inefficient; and tracking action items across meetings and compiling briefs rely on memory, leading to discontinuity.

How the Skill Works: Core Mechanisms and Key Steps

This skill uses a four-tier routing system to handle varying complexities:
- A·Quick: Directly answers queries about meeting types or history without generating files.
- B·Mini: Processes text ≤2000 characters, executing the full workflow.
- C·Full: Handles ≥2000 characters, compliance decisions, or archival needs via a 10-step pipeline.
- B'·Brief Tier: Aggregates multiple minutes into a monthly briefing docx.

For transcripts, the skill automatically runs transcription proofreading (§2.5). Triggered by timestamps or speaker labels, it applies a three-tier confidence approach: high-confidence corrections (e.g., from a known-name table) are applied directly, medium-confidence issues are marked [?to-verify], and numbers/names are flagged only without modification. Proofreading follows three non-negotiable rules and leverages references/asr-纠错对照表.md for continuous learning.

Minutes generation follows a four-step workflow:
1. Context: Identifies the client, meeting type, and retrieves historical slices.
2. Execute: First builds an evidence matrix (argument/evidence type/source/position), then drafts content with every factual sentence tagged as [实证] (evidence-based) or [估算] (estimated).
3. Self-check: Validates sentence-level tri-state coverage, four mandatory action item fields (content/owner/deadline/deliverable), and client anonymization.
4. Deliver: Outputs a docx file (per 附件10 template), markdown full text, and a memory slice.

For recurring meetings (e.g., weekly standups), the skill activates the B·Weekly Tier, which automatically extracts previous action items, verifies project entities (e.g., client names/amounts), and generates both minutes docx and a tracking xlsx with week-over-week action linkage.

Usage Boundaries and Important Considerations

The skill has defined constraints:
- Input: Only supports .docx, .txt, or .md text files; audio/video is rejected—must be ASR-transcribed first.
- Auto-trigger: Proofreading runs automatically when ASR sources are detected, but ambiguous inputs (e.g., unclear tier) prompt an AskUserQuestion without assuming a tier.
- Human review required: The briefing tier (B') mandates user approval before outputting docx.
- Special cases: One-on-one interviews hand off to the interview-insight skill; cross-meeting inquiries (e.g., priority or deliverable queries) scan slices from the last 90 days with weighted ranking.
- Memory system: All minutes and briefs are stored under memory/projects/clients/ for cross-session use, but client codenames must be registered first (§6).

Note: Action items in minutes must include four elements, with deadlines avoiding vague terms like "近期" (soon) or "尽快" (ASAP); briefs use a FangSong+HeiTi font scheme, distinct from minutes' Microsoft YaHei.