Longform Synthesis

Map-reduce long documents into cited syntheses.

Skill metadata

Source	Bundled (installed by default)
Path	skills/research/longform-synthesis
Version	1.0.0
Author	VibeOS
License	MIT
Platforms	linux, macos, windows
Tags	research, analysis, summarization, long-context, synthesis
Related skills	professional-research, source-critique, academic-literature

Reference: full SKILL.md

信息

The following is the complete skill definition that VibeOS loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.

Longform Synthesis Skill

Turn long sources (reports, threads, PDFs, multi-URL extracts) into a tight synthesis without inventing quotes. Preserve numbers and named claims.

When to Use

• Source text is too large for one careful pass (multi-chapter, many URLs).
• User wants an executive synthesis with citations back to chunks/URLs.
• After web_extract returns summaries that lost structure you still need.

Prerequisites

• Sources as files and/or URLs already fetched.
• Prefer read_file / web_extract / browser snapshot over guessing.
• Do not swap toolsets mid-session; work with what is loaded.

How to Run (map → reduce)

1. Inventory — list sources with approximate size and role.
2. Chunk — split each source into coherent sections (headings, pages, or ~3–8k char blocks). Keep chunk ids stable (S1-C3).
3. Map — for each chunk, extract:
   • claims (factual)
   • numbers / dates
   • definitions
   • open questions Tag each bullet with chunk id.
4. Reduce — merge maps; dedupe; cluster themes; note conflicts.
5. Write — executive synthesis + evidence table citing chunk ids / URLs.
6. Verify — every non-obvious claim has a citation; conflicts listed.

Helper (optional) — from this skill directory:

python3 scripts/chunk_text.py --max-chars 6000 < long_doc.txt > chunks.json

Output template

# Synthesis: <topic>

## Executive summary
...

## Themes
### Theme A
- Finding … [S1-C2]
- Finding … [S2-C1]

## Numbers & dates
| Value | Context | Cite |
|-------|---------|------|

## Contradictions
- …

## Gaps
- …

## Sources / chunks
- S1: URL or path
  - C1: …

Pitfalls

• Do not paraphrase away digits, legal names, or version numbers.
• If web_extract summarized away fields you need, re-fetch with browser tools.
• Stay cache-safe: no mid-turn toolset changes.
• Cap map fan-out — prefer sequential maps over dozens of parallel dumps.

Verification

• Every theme bullet has a chunk/URL cite.
• Contradictions section exists (or “none found”).
• Executive summary ≤ ~12 sentences unless user asked for more.