dictation

Scribe: Dictation - Idea Organizer

Parse unstructured input (speech-to-text, typed ramblings, braindumps), identify distinct topics, merge into existing notes or create new ones, and maintain a searchable index. Input may be messy, fragmented, or imperfect — that's expected.

Output Directory

Notes go into claude-notes/ in the current project directory to avoid collision with existing project files.

project-root/
|- claude-notes/
|  |- INDEX.md
|  |- topic-one.md
|  \- ...
|- existing-project-files/
\- ...

No exceptions. Always use claude-notes/ in the current working directory, even outside a git repo. Never use ~/.claude/notes/.

First run: If claude-notes/ or claude-notes/INDEX.md does not exist, create them.

Workflow

0. Load or Initialize the Working User Profile

Use ~/.scribe/user-profile.md as the active profile for personalization.

• If ~/.scribe/user-profile.md exists, read it first.
• If missing, create a neutral scaffold with # Communication Profile: User, Last Updated, Sessions Analyzed: 0, **Profile Format:** v2, and the standard profile sections (including the Candidate Patterns table).
• Use references/user-profile.md only as a public example for section quality and pattern granularity.
• If you copy anything from the public example, place it under ## Candidate Patterns (Need More Evidence) with a [seed-example] marker, not as confirmed behavior.
• Never overwrite references/user-profile.md during normal runs; treat it as the public baseline example.
• v1 migration: If the profile exists but lacks **Profile Format:** v2, migrate it per the v1→v2 procedure in shared/references/session-tracking.md.

0b. Reconciliation Pass

Run the reconciliation procedure from shared/references/session-tracking.md:

1. Glob ~/.scribe/sessions/*.md.
2. For each candidate in the profile's Candidate Patterns table, grep session files for its slug and recount.
3. If any candidate now has 2+ sessions of evidence, promote it to the appropriate main section.
4. Update the profile if any changes were made.

This catches promotions lost to concurrent writes or interrupted sessions.

1. Parse the Dictation

• Read raw text carefully.
• Identify distinct ideas or topics (may be 1 or many).
• Separate unrelated thoughts into individual note candidates.
• Key insight extraction: Users often bury the core idea partway through. Scan for transition phrases: "so what would be useful is...", "the idea is...", "it'd be great if..."
• Priority from tone:
  • High: excitement, detailed elaboration, connecting to multiple concepts.
  • Low: "probably not worth exploring", hedging.
  • Meta/feedback: "I like how you're doing X" - act on it, do not create a note.

2. Build Idea Landscape and Check Existing Notes

Read claude-notes/INDEX.md. Reason about relationships between existing notes.

Merge heuristic:

• Merge when new content extends, refines, or adds a dimension to an existing note.
• Create new when content is conceptually independent even if it shares tags.
• Ask: "Does this make the existing note more complete, or does it stand on its own?"

3. Write or Update Notes

See references/note-schema.md for the full format. Mandatory quality standards:

• Every section filled in - no empty or placeholder sections.
• Idea section distills dictation into clear prose - not just echoed speech.
• Related links verified (files exist, cross-references are real).
• Next Steps are concrete and actionable.
• Timestamps on creation and every update.

4. Update the Index

Always update claude-notes/INDEX.md after every change. Each entry includes:

File	Summary	Tags	Source	Created	Updated

5. Learn and Update the Working User Profile

Use the two-phase session tracking protocol from shared/references/session-tracking.md:

Phase 1 — Write Session Record:

• Create ~/.scribe/sessions/ if it doesn't exist.
• Write a session file named {YYYYMMDDTHHMMSS}-dictation-{5-random}.md.
• Record observed patterns (tagged with [idea-structure], [priority-signal], [style-pref], [behavior]), explicit statements, and candidate evidence with canonical slugs.
• Session files are write-once, never modified.

Phase 2 — Update Profile with Evidence Counting:

• For each observed pattern, determine its canonical slug.
• Grep across all files in ~/.scribe/sessions/ for each slug to count distinct sessions.
• Apply promotion rules:
  • Explicit statement in this session → promote immediately to the appropriate main section.
  • 2+ sessions containing the slug → promote from Candidate Patterns to the appropriate main section.
  • <2 sessions → add or update the row in the Candidate Patterns table (update Sessions count, Last Seen date).
• Update Last Updated, Sessions Analyzed, and Profile Evolution Log.
• Keep profile updates additive; do not delete prior observations.
• Best-effort git commit of session file + profile to ~/.scribe repo.

6. Respond

Brief summary:

• Topics identified (count).
• Notes created vs. updated (with filenames).
• Cross-references found.
• Priority level detected.
• Profile updates applied (or none).
• Ambiguous points needing clarification.

Merge Rules

• Additive only - never delete or overwrite existing note content.
• New content gets a fresh timestamp.
• If entries contradict, keep both with timestamps - temporal record matters.
• Append raw dictation to the Source Log.

File Naming

• Kebab-case: threat-model-retrospective.md
• Descriptive but concise.
• No date prefixes - timestamps live inside the file.

Quick Invocation

• /scribe:dictation or "take dictation"
• Or just paste raw text