Survey Report

Goal: Answer the question the user asked from a survey CSV with facts, and save a short report as Markdown to reports/.

The user's question is the request they made when invoking this skill. If it contains a file path/name, use that CSV; otherwise use the CSV in the current folder.

Tool

All numbers come from the shared engine — never estimate or count "in your head". Standard library only, no pandas required.

python3 scripts/survey.py <cmd> [...] [--file CSV]

Run these from the project root. If scripts/survey.py is not there, check .claude/scripts/survey.py or locate survey.py in the project.

• profile — overview: n, every column with index, detected type (META / SINGLE / MULTI / FREETEXT), fill rate. Always run first.
• freq COL [--by COL2] [--filter "COL=VALUE"] [--top N] — frequency distribution; --by = cross-tab; --filter repeatable (subset). For MULTI columns the percentages sum to > 100% (correct — explain it in the report).
• text COL [--filter ...] [--sample N] — free-text answers for qualitative themes.

COL = column index (from profile) or an unambiguous part of the column name. Every command accepts --json for machine-readable output.

Survey context

If a context document sits next to the CSV (<name>.context.md, or survey-context.md when the folder holds exactly one CSV), read it before writing. It carries the subject, the glossary, the audience and the study design. Use it for framing and terminology — the figures still come only from the engine. Whatever in it limits the findings, above all the "Who is missing" section, belongs in this report's interpretation and method. Any expectations listed there are something to test, never something to support: never cite one as evidence and never let one shape how a finding is worded. If no such document exists, analyze as before and mention survey-context once at the end of your reply.

Workflow

1. Profile. Run profile to learn n, column indices and types.
2. Translate the question into analyses. Pick the right column(s) and run the fitting commands:
   • "how many / what share / distribution" → freq on the target column.
   • "does X differ by Y / relationship / comparison" → freq X --by Y.
   • "… for group Z" → --filter "Z-column=value".
   • "why / what do they value / typical workflow / wishes" → text on a free-text column; with many answers use --sample 40–60, cluster themes, back them with ~2–3 short verbatim quotes. Usually 2–4 analysis commands in addition to profile — a bare distribution is rarely the whole answer. For open questions combine angles (distribution + relevant cross-tab + free-text themes).
3. Write the report (see format); numbers exactly as the engine returned them.
4. Save to reports/. Create the folder if needed. File name: YYYY-MM-DD_short-title.md (date via date +%F, title kebab-case from the question). If a file already exists, append -2 etc.; never overwrite.
5. Tell the user the key takeaway briefly in your reply and give the report path.

Report format (keep it short)

# <question as a concise heading>

**Data basis:** <file> · n = <responses> · report of <date>

## Key takeaway
1–3 sentences answering the question directly.

## Results
- Supported points with **numbers (absolute + %)**. Tables for distributions/cross-tabs.
- For multi-select, note: multiple answers possible, sum > 100%.

## Interpretation
What does it mean? Notable patterns, relationships, caveats (e.g. low fill rate of a
column, small subset after filtering, limits of the study design from the survey context).

## Method
Columns and filters used, in one sentence, so it's reproducible. If a context document was
used, name the recruiting route and who it cannot reach.

Rules

• Accuracy over volume. Only claim what the outputs support. Percentages relate to the respondents who answered the column (the engine states this).
• Compute aggregates ("positive overall", top-2-box) from the absolute counts, never by adding the engine's rounded percentages: 73 of 140 is 52.1%, not 13.6% + 38.6%.
• Statistics in plain text (Chi² = 12.094, p = 0.438, Cramér's V = 0.17), no LaTeX. Pass on the engine's reliability warnings rather than dropping them.
• Flag small filtered subsets (n < ~30) as weakly reliable.
• Report language = language of the question (default English).
• No personal raw data (e.g. individual ids/timestamps).
• Keep it a short report, not a full study.