科研 Skill

QualiGPT主题分析

qualigpt

Runs QualiGPT-style batch thematic analysis across interviews or documents, producing a theme table with quotes and source counts. Use for LLM-assisted thematic coding of a document batch.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

QualiGPT: LLM-Assisted Qualitative Thematic Analysis

This skill replicates the methodology of QualiGPT (Zhang et al., https://github.com/KindOPSTAR/QualiGPT — cite arXiv:2310.07061 and arXiv:2309.10771), a tool for LLM-assisted thematic analysis of qualitative data. The original tool sends structured prompts to an LLM; here, Claude performs the analysis directly using the same prompt structure, segmentation strategy, and output format.

Methodology overview

QualiGPT automates the coding step of qualitative analysis: identifying key themes in a dataset, describing them, and grounding each theme in verbatim quotes with a count of how many participants/sources support it. It follows this pipeline:

  1. Ingest data (.docx, .xlsx, .csv, .txt). For tabular data, expect a header row like Speaker | Content.
  2. Determine dataset type — interview, focus group, social media posts, or archival documents. Infer from context if obvious (e.g., the user says "these are my interview transcripts"); otherwise ask.
  3. Specify number of themes N (QualiGPT default: 10, range 1–20). Ask the user or default to 10.
  4. Ask the user to choose coding mode before analyzing, unless they've already made it clear: inductive (themes emerge from data — default if they have no codebook or don't state a preference) or deductive (they supply a codebook; themes must map to it). Skip the question only if the user has already handed you a codebook, named specific themes to code for, or otherwise unambiguously signaled deductive intent — don't ask a redundant question, but don't silently assume inductive when a codebook or a preference is on the table.
  5. Segment if the corpus exceeds the context that can be carefully analyzed in one pass (original: ~3,800 tokens per segment, split on sentence boundaries, never mid-sentence). Analyze each segment separately.
  6. Consolidate: if multiple segments were analyzed, merge all segment-level theme tables and re-analyze the merged responses to produce one final consolidated theme table (deduplicate overlapping themes, re-tally counts).
  7. Export results as .csv (parse the pipe-delimited table) and a human-readable report.

Analysis prompt (apply this to each segment)

Act as an experienced qualitative researcher. Apply this instruction pattern, adapted from QualiGPT verbatim:

You need to analyze a dataset of {interviews | a focus group | Social Media Posts | archival documents}. Please identify the top {N} key themes from the data and organize the results in a structured table. The table must have exactly 4 columns: Theme, Description, Quotes, and Participant Count (for documents: Source Count — number of distinct documents supporting the theme). Use | to separate columns, include a header separator row |---|---|---|---|, and end the table with a line of ********** before and after so it can be parsed.

Rules for faithful application:

  • Quotes must be verbatim from the source text — never paraphrase inside the Quotes column. Include 1–3 short representative quotes per theme, each attributed to its source document/speaker.
  • Counts must be real tallies — count the distinct participants/sources that evidence the theme; do not estimate.
  • Themes should be distinct (minimal overlap), descriptive, and ranked roughly by prevalence.
  • For deductive coding, only use themes from the user's codebook; note codebook themes with no supporting data.

Batch workflow for multiple files

  1. Convert every file to plain text (e.g., textutil -convert txt for .docx on macOS; pandas for .xlsx/.csv). Keep a file → source ID mapping.
  2. Read every document in full. Prepend each document's text with a source marker like [SOURCE: doc_id] so quotes can be attributed.
  3. If the total corpus is small enough to analyze carefully in one pass (roughly < 8,000 words), analyze the whole corpus at once. Otherwise segment per the rules above, analyze each segment, then run the consolidation pass.
  4. Produce three deliverables in an output folder next to the source files (e.g., qualigpt_output/):
    • themes.csv — the final theme table parsed into CSV (columns: Theme, Description, Quotes, Source Count, Source IDs)
    • analysis_report.md — the formatted theme table plus a short narrative summary, methodology note, and per-document inventory
    • raw_response.txt — the full pipe-delimited table(s) with ********** markers, mirroring QualiGPT's raw output format
  5. Report the top themes to the user in chat with the table.

Feeding a codebook (bridge to beebe-policy-framing)

An inductive qualigpt run is the recommended first stage before deductive analysis with the beebe-policy-framing skill: its theme table is the raw material for that skill's codebook (themes → frame families; attributed speakers → actor codes; contested programs → activity codes). When the user's goal is framing/metaphor analysis, offer this two-stage path rather than asking them to invent a codebook cold.

Limitations to disclose

Per the QualiGPT paper: LLM-assisted coding is a first-pass aid, not a replacement for researcher judgment. Recommend the researcher (a) verify quotes against sources, (b) review theme boundaries, and (c) treat counts as approximate guides for salience. Note any documents that contributed little signal (e.g., forms, boilerplate). If documents duplicate one another, flag it — duplicates inflate source counts.

示例提示词

帮我对这批焦点团体访谈记录批量做主题分析并产出主题表。

所属插件

社会学、政治学与城市研究面向社会科学研究:问卷与访谈设计、质性编码、文本即数据、社会网络、ABM、比较政治与选举数据、RCT 与 DID、城市与区域分析。

安装该插件会一并获得同领域的 58 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

QualiGPT主题分析能做什么?

Runs QualiGPT-style batch thematic analysis across interviews or documents, producing a theme table with quotes and source counts. Use for LLM-assisted thematic coding of a document batch。

如何安装QualiGPT主题分析?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「QualiGPT主题分析」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

QualiGPT主题分析需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

QualiGPT主题分析经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。