# Survey Seed Harvest

Identify survey/review papers in a retrieved set and extract taxonomy seeds into...

- Repository: https://github.com/WILLOSCAR/research-units-pipeline-skills (517 stars)
- Category: Engineering
- License: see repository
- Last updated: 2026-10-05
- Runs scripts: yes, review before use

## Install

Claude Code: `npx skills add WILLOSCAR/research-units-pipeline-skills --skill survey-seed-harvest -g -a claude-code -y`
Codex: `npx skills add WILLOSCAR/research-units-pipeline-skills --skill survey-seed-harvest -g -a codex -y`
Cursor: `npx skills add WILLOSCAR/research-units-pipeline-skills --skill survey-seed-harvest -g -a cursor -y`
Gemini CLI: `npx skills add WILLOSCAR/research-units-pipeline-skills --skill survey-seed-harvest -g -a gemini-cli -y`

## Description

Identify survey/review papers in a retrieved set and extract taxonomy seeds into `outline/taxonomy.yml` (topics/subtopics/terminology).

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# Survey Seed Harvest

## Triggers & routing

- **Trigger**: survey seed harvest, taxonomy seeds, 从 survey 提 taxonomy, bootstrap taxonomy.
- **Use when**: retrieval/dedup 后想快速从已有 survey/review 论文中提取术语与主题结构，用于加速 `taxonomy-builder`。


Bootstrap taxonomy *seeds* from existing survey/review papers inside your retrieved set.

This is an accelerator for the early structure stage: it should make `taxonomy-builder` easier, not replace it.

## Inputs

- `papers/papers_dedup.jsonl` (deduped paper metadata with titles/abstracts)

## Outputs

- `outline/taxonomy.yml` (seed taxonomy; expected to be refined)

## Workflow (heuristic)
Uses: `papers/papers_dedup.jsonl`.


1. Find likely survey/review papers:
   - title/abstract contains “survey”, “review”, “systematic”, “meta-analysis”
2. Extract candidate topic terms and group them into:
   - ~4, 10 top-level nodes (“chapters”)
   - 2, 6 children per node (mappable leaves)
3. Write short, *actionable* descriptions:
   - what belongs here / what does not
   - (optional) list 2, 5 representative titles as seeds
4. Treat the result as a starting point:
   - pass it to `taxonomy-builder` for domain-meaningful rewriting and scope alignment.

## Quality checklist

- [ ] `outline/taxonomy.yml` exists and is valid YAML.
- [ ] Taxonomy has at least 2 levels (`children` used) and every node has a description.
- [ ] Avoid generic placeholder nodes like “Overview/Benchmarks/Open Problems” unless they are truly content-based for your domain.

## Script (optional helper)

### Quick Start

- `uv run python .codex/skills/survey-seed-harvest/scripts/run.py --help`
- `uv run python .codex/skills/survey-seed-harvest/scripts/run.py --workspace <workspace>`

### All Options

- `--top-k <n>`: number of candidate terms to consider
- `--min-freq <n>`: minimum frequency threshold

### Examples

- More conservative term selection:
  - `uv run python .codex/skills/survey-seed-harvest/scripts/run.py --workspace <workspace> --top-k 80 --min-freq 3`

### Notes

The rest: https://github.com/WILLOSCAR/research-units-pipeline-skills/blob/HEAD/.codex/skills/survey-seed-harvest/SKILL.md