Dovetail

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the research repository I actually need instead of Dovetail.

The product is a database of quotes with tags on them. What makes it valuable is discipline: **every claim traceable to the sentences that produced it.** Build that and the rest is a list.

STACK
- Node 20+ with Fastify, server-rendered HTML with a little vanilla JS
- SQLite through better-sqlite3, WAL mode, FTS5 over transcripts
- whisper.cpp for transcription
- Caddy in front

THE DATA MODEL
- studies: id, name, question, method, started_at
- sessions: id, study_id, participant_ref, held_at, media_path, transcript_status
- segments: id, session_id, speaker, start_ms, end_ms, text
- highlights: id, segment_id, start_char, end_char, text, created_by, created_at — **the quote, exactly as said**
- tags: id, name, group, description, is_deprecated
- highlight_tags: highlight_id, tag_id, applied_by, at
- themes: id, name, statement, confidence, study_id
- theme_highlights: theme_id, highlight_id — a theme is a claim plus its evidence and nothing else
- participants: id, pseudonym, segment_attrs_json — **no names, ever**

GETTING INTERVIEWS IN
- Upload audio or video, transcribe locally with whisper.cpp, diarise into speakers
- The transcript is the working surface: select text, press a key, it becomes a highlight
- Import an existing transcript as text, because half the material always arrives that way

TAGGING WITHOUT THE TAXONOMY ROTTING
- Tags grouped — behaviour, pain, desire, context — and every tag carries a one-line definition. **An undefined tag means two researchers use it differently and the archive stops being comparable**
- Merging tags moves every application and records the merge
- A deprecation flag rather than deletion, so old work is still readable
- Suggest existing tags as somebody types; that alone halves the duplicate rate

THEMES, WHICH ARE THE OUTPUT
- A theme is a sentence somebody is prepared to defend, with the highlights attached
- Show the count of distinct participants, not the count of highlights. **Five quotes from one talkative person is not a pattern**, and this single number prevents the most common research error
- Contradicting evidence attachable to the same theme, and shown. A repository that only stores confirming quotes is a machine for confirmation bias

PRIVACY, BECAUSE THIS IS INTERVIEW DATA
- Pseudonyms in the interface; the mapping to real people kept separately with narrow access
- Redaction on the transcript that also cuts the audio, not just the text
- Retention per study with a sweeper, consent recorded per session, and deletion on request that reaches the media

SHARING
- A finding renders as the statement, the participant count, and the quotes with play buttons
- Read-only links for the rest of the company, which is how research stops being a document nobody opens

WHAT MATTERS MOST
The distinct-participant count on every theme, and defined tags. One keeps the conclusions honest; the other keeps the archive usable after the third researcher joins.