prompt-distiller

Heuristic prompt-pattern miner. Scans a folder of .txt/.md prompts (or a single conversation export) and emits a prompts.yaml of reusable templates with {placeholder} variables, plus a frequency report.

No LLM call. Pure Python stdlib + regex. Token-frequency, regex shape detection, and line-level diff to find the variable bit between near-duplicates.

Usage

python3 scripts/distill.py --input <folder-or-file> [--output prompts.yaml] [--top 10] [--min-uses 3]

Flags

• --input — folder of .txt/.md, or a single file (conversation export). Required.
• --output — path to write the templates YAML. Default prompts.yaml.
• --top — emit at most N templates (highest-frequency first). Default 10.
• --min-uses — minimum occurrences for a cluster to become a template. Default 3.

How it works

1. Parse all input texts. Split into "prompts" by:
   • blank-line-separated paragraphs, or
   • lines starting with human: / user: / you: (chat exports).
2. Normalize whitespace and collect token sequences.
3. Find recurring 3-7 word openers (e.g. "write me a", "fix the bug in").
4. Bucket prompts by fuzzy structure (token-edit similarity). Within each bucket, line-diff identifies the variable spans and replaces them with {var1}, {var2}, ...
5. Rank by uses. Emit YAML + stdout summary.

Output

templates:
  - id: bug-fix
    body: "Fix the bug in {file} where {issue}"
    uses: 14
    examples: ["Fix the bug in agent.js where it loops forever", ...]
  - id: explain
    body: "Explain how {thing} works"
    uses: 9

When to use

• You have a graveyard of past prompts and want a reusable library.
• You want to know what you keep typing over and over.
• You want to seed snippet/expander tools (TextExpander, Raycast, etc).