wbern · May 14, 2026 16:13
diff --git a/gistfile1.txt b/gistfile1.txt
 Make a word-cloud of my favorite short Claude Code prompts.

 1. Scan every `.jsonl` under `~/.claude/projects/`. Each line is one transcript event.
 2. Keep only events with `type:"user"` whose `message.content` is a **string** (skip tool results, which are lists). Also skip events with `isCompactSummary:true` or `isVisibleInTranscriptOnly:true`.
 3. Drop content that isn't a user keystroke:
   - anything starting with `<` followed by a lowercase tag — e.g. `<bash-input>`, `<command-name>`, `<command-message>`, `<command-args>`, `<command-stdout>`, `<command-stderr>`, `<local-command-stdout>`, `<bash-stdout>`, `<bash-stderr>`, `<system-reminder>`, `<user-memory-input>`, `<ide_*>`
   - anything starting with `/` (slash commands like `/clear`)
   - system echoes: `[Request interrupted ...]`, `unknown command: ...`
 4. Normalize: lowercase, collapse whitespace, strip trailing `.,;:!?`. Track the `sessionId` for each occurrence.
 5. Within each transcript file, sort events by timestamp and collapse runs of the same normalized phrase to one occurrence — repeated identical prompts in a row are duplicate submits.
 6. Keep prompts that are **≥ 2 characters and ≤ 10 words**. Drop single-character replies (`y`, `1`, `a`). Count exact-phrase occurrences across all transcripts. **Drop singletons (keep count ≥ 2).**
 7. **Cross-session uniformity filter** — humans paraphrase, daemons template: for any phrase with **≥ 3 words**, count the number of distinct `sessionId`s it appears in. If that exceeds **5**, drop it as likely daemon/hook injection.
 8. **Final structural-noise pass** — drop only entries that match a concrete automation shape:
   - `ALL_CAPS_TOKEN:` prefix (e.g. `HEALTH_CHECK:`, `URGENT:`)
   - `key=value` fragments (e.g. `exit=completed`, `pid=1234`)
   - contain explicit status verbs in templated position: `dispatched`, `slung`, `enqueued`, `exit=`

   Do **NOT** drop entries just because they are high-frequency, terse, or feel unusual. A human typing `address` 84 times is plausible. When in doubt, keep.
 9. Render an SVG word cloud — each phrase once, font size ∝ frequency, largest centred, non-overlapping spiral placement. Save to `./prompt_cloud.svg` and open it.
 10. Print the top 20 phrases with counts as a markdown table, and briefly list any phrases you dropped in steps 7 and 8 with their counts so I can sanity-check.
	Make a word-cloud of my favorite short Claude Code prompts.

	1. Scan every `.jsonl` under `~/.claude/projects/`. Each line is one transcript event.
	2. Keep only events with `type:"user"` whose `message.content` is a string (skip tool results, which are lists). Also skip events with `isCompactSummary:true` or `isVisibleInTranscriptOnly:true`.
	3. Drop content that isn't a user keystroke:
	- anything starting with `<` followed by a lowercase tag — e.g. `<bash-input>`, `<command-name>`, `<command-message>`, `<command-args>`, `<command-stdout>`, `<command-stderr>`, `<local-command-stdout>`, `<bash-stdout>`, `<bash-stderr>`, `<system-reminder>`, `<user-memory-input>`, `<ide_*>`
	- anything starting with `/` (slash commands like `/clear`)
	- system echoes: `[Request interrupted ...]`, `unknown command: ...`
	4. Normalize: lowercase, collapse whitespace, strip trailing `.,;:!?`. Track the `sessionId` for each occurrence.
	5. Within each transcript file, sort events by timestamp and collapse runs of the same normalized phrase to one occurrence — repeated identical prompts in a row are duplicate submits.
	6. Keep prompts that are ≥ 2 characters and ≤ 10 words. Drop single-character replies (`y`, `1`, `a`). Count exact-phrase occurrences across all transcripts. Drop singletons (keep count ≥ 2).
	7. Cross-session uniformity filter — humans paraphrase, daemons template: for any phrase with ≥ 3 words, count the number of distinct `sessionId`s it appears in. If that exceeds 5, drop it as likely daemon/hook injection.
	8. Final structural-noise pass — drop only entries that match a concrete automation shape:
	- `ALL_CAPS_TOKEN:` prefix (e.g. `HEALTH_CHECK:`, `URGENT:`)
	- `key=value` fragments (e.g. `exit=completed`, `pid=1234`)
	- contain explicit status verbs in templated position: `dispatched`, `slung`, `enqueued`, `exit=`

	Do NOT drop entries just because they are high-frequency, terse, or feel unusual. A human typing `address` 84 times is plausible. When in doubt, keep.
	9. Render an SVG word cloud — each phrase once, font size ∝ frequency, largest centred, non-overlapping spiral placement. Save to `./prompt_cloud.svg` and open it.
	10. Print the top 20 phrases with counts as a markdown table, and briefly list any phrases you dropped in steps 7 and 8 with their counts so I can sanity-check.
No results found