prompt-archaeology

name

prompt-archaeology

description

Recover lost prompts, spoken ideas, and partially submitted instructions from local AI session logs, voice transcription logs, app-focus telemetry, prompt-history files, and pasted-text artifacts. Use when the user asks to find a prompt they remember saying or typing, especially with fuzzy keywords, approximate dates, Always/Superwhisper/voice dictation, Codex/Claude/Devin sessions, or questions like "where did my prompt/code go?"

Prompt Archaeology

Overview

Reconstruct missing prompts by treating remembered phrases as anchors, then expanding through nearby timestamps, focused apps, pasted text, and AI session logs. Prefer evidence over recall: report exact files, lines, timestamps, and whether the prompt was spoken, pasted, or actually submitted.

Workflow

1. Start with the user's strongest anchor words.

   • Search exact variants first: original wording, singular/plural, grammar-corrected spelling, and likely transcription mistakes.
   • For voice prompts, search concept clusters too: related nouns, tools, model names, provider names, and distinctive phrases.

2. Bound the time window.

   • Use the user's date memory if available.
   • Search by month/day first rather than global broad searches.
   • If a broad search is needed, exclude noisy directories such as node_modules, browser caches, and generated build outputs.

3. Search in layers.

   • AI transcripts: ~/.codex/sessions, ~/.claude, Devin local transcripts, project .jsonl logs, prompt-history state files.
   • Voice logs: ~/Library/Logs/Always/always.YYYY-MM-DD, ~/Library/Application Support/superwhisper, ~/Library/Application Support/vox.
   • Draft/paste artifacts: Codex attachments, pasted-text.txt, editor histories, prompt drafts.
   • App telemetry: Always uds_focused_app_changed, transcription_received, transcription_pasted, and submit/abort events.

4. When an anchor hits in a voice log, reconstruct the speech window.

   • Extract all transcription_received rows around the anchor, usually 5-15 minutes before and after.
   • Also extract transcription_pasted and uds_focused_app_changed in the same window.
   • Use transcription_received as raw evidence and transcription_pasted/grammar-corrected text as a cleaned version.
   • Watch for focus changes that explain where the prompt was drafted or submitted.

5. Verify whether the prompt was submitted.

   • A voice log proves it was spoken/pasted, not necessarily sent.
   • Look for app focus moving to ChatGPT/Codex/Claude/Devin right after the pasted sequence.
   • Search exact cleaned phrases in AI session logs after the voice timestamp.
   • State the confidence level: submitted prompt, pasted draft, spoken-only idea, or reconstructed prompt.

6. Present the result compactly.

   • Lead with the likely answer and why.
   • Include clickable absolute file links with line numbers.
   • Give a reconstructed prompt only after naming the raw source and timestamps.
   • Do not print secrets or API keys found in logs; redact or summarize credential evidence.

Useful Commands

Extract a focused Always window:

scripts/extract-always-window.sh ~/Library/Logs/Always/always.2026-06-13 2026-06-13T00:47:00Z 2026-06-13T00:58:00Z

Search June Always speech for conceptual anchors:

for f in ~/Library/Logs/Always/always.2026-06-*; do
  jq -r 'select(.fields.message == "transcription_received")
    | select(.fields.text | test("determin|provider|providers|model|models|harness|TPS|token per second|visual|OpenRouter|static JSON"; "i"))
    | [.timestamp, .fields.text] | @tsv' "$f" 2>/dev/null
done

Search AI session logs without drowning in project files:

rg -n -i --hidden --glob '!node_modules/**' --glob '!Library/**' \
  'exact phrase|anchor phrase|distinctive tool name' \
  ~/.codex ~/.claude ~/.local/share/devin ~/Documents 2>/dev/null

Query Superwhisper SQLite by date and anchors:

sqlite3 -tabs "$HOME/Library/Application Support/superwhisper/database/superwhisper.sqlite" \
  "SELECT r.datetime, coalesce(c.c3,c.c2,c.c1,'')
   FROM recording r JOIN recording_fts_content c ON c.c0 = r.id
   WHERE r.datetime >= '2026-06-01'
     AND r.datetime < '2026-06-20'
     AND lower(coalesce(c.c3,c.c2,c.c1,'')) LIKE '%harness%'
   ORDER BY r.datetime;"