Shama Add Context

Add Shama Context

Add new entries to Shama's vocabulary, context, or instructions files when the transcriber misrecognizes words, the cleanup step formats text incorrectly, or shell-mode generates the wrong command.

Shama is the resident macOS voice-to-text app. It reads a per-project voice.json plus an optional global voice.json; the file keys below point at the markdown files that feed its pipelines.

Two context levels: project-level (default) — files specific to the current project, stored alongside the project's voice.json — and global-level — files shared across all projects, stored alongside ~/.voice/voice.json. For each file key, Shama takes the union of global then project values: global file-list entries come first, project entries are appended (duplicates by resolved path removed). A project-level entry adds to — it does not replace — the global one.

Two modes, each with its own trio of files:

• Clipboard mode (Control+Option+Space) — contextFile, vocabularyFile, instructionsFile
• Shell-command mode (Control+Option+S) — shellContextFile, shellVocabularyFile, shellInstructionsFile

Each pipeline consumes different files:

• Transcription step (OpenAI) — the relevant vocabulary file: a flat list of terms sent as hints for accurate word recognition. Affects what words the transcriber produces.
• Cleanup step (Claude, clipboard mode) — contextFile (project knowledge) and instructionsFile (how-to formatting rules). Affects how transcribed text is corrected and structured.
• Shell-command generation (Claude, shell mode) — shellContextFile (background for command generation) and shellInstructionsFile (rules for command output). Affects what shell command is produced.

Identify which stage is failing:

• Wrong word recognized → vocabulary (or shell-vocabulary)
• Right word, wrong formatting / structure → context or instructions
• Shell command wrong style, tool choice, or OS flags → shell-context or shell-instructions

Step 1: Determine the target level and locate the files

Use global-level if the user explicitly mentions "global", "global context", "global config", "global Shama context", or ~/.voice/. Otherwise default to project-level.

Global-level resolution

1. Read ~/.voice/voice.json to find any of the six file keys above
2. Resolve all configured paths relative to ~/.voice/
3. If ~/.voice/voice.json does not exist or has none of these keys configured, inform the user that global Shama config has no context files yet and offer to create one — see "Creating a missing file" in Step 3
4. Read the contents of any files that exist

Project-level resolution (default)

1. Read voice.json in the current directory to find any of the six file keys
2. If voice.json does not exist, check for the conventional filenames in the current directory:
   • voice-context.md, voice-vocabulary.md, voice-instructions.md
   • voice-shell-context.md, voice-shell-vocabulary.md, voice-shell-instructions.md
3. If no project files exist at all, inform the user that Shama is not configured for this project and suggest running /shama-init-config first. Stop here.
4. Read the contents of any files that exist

Step 2: Determine the mode and target file(s)

Default to clipboard mode. Use shell-command mode when the user mentions "shell", "shell mode", "shell command", "terminal", "Control+Option+S", or shell-mode dictation.

If the user's request is too vague to act on, ask for:

• The word or phrase as it was incorrectly transcribed (or the wrong shell command)
• The correct form it should produce
• Optionally, a brief definition or context note

Then pick the file(s) for the scenario:

Clipboard mode (Control+Option+Space)

Shell-command mode (Control+Option+S)

Step 3: Add the entries

Vocabulary file entries (clipboard or shell)

One term per line — no markdown formatting, no descriptions, just the bare term with correct spelling and capitalization:

Zod
CLAUDE.md
ResolvedFileRef

Context file entries (clipboard mode)

Add to the appropriate section, following the existing format:

• Technologies: - {Name} ({brief description})
• Terminology: - **{Term}** - {brief definition or correction note}
• Naming Conventions: - {convention description}

When adding a misheard word correction, include the common misheard form in the definition to help Claude recognize and correct it:

- **Zod** - Runtime validation library (not "god", "sod", or "zawed")
- **lgit** - Git wrapper for .local directory (not "legit" or "l-git")

Instructions file entries (clipboard mode)

The instructions file controls how cleanup is performed. Add formatting rules, style preferences, or behavioral overrides — not project knowledge:

- Always preserve fenced code blocks verbatim — do not edit, reformat, or summarize their contents
- Prefer hyphen-style bullet points (`-`) over asterisks (`*`)
- When listing steps, use numbered lists; when listing options, use bullets

Shell-context file entries (shell mode)

Background information for the shell-command generator: project paths, repo layout, preferred tools, common workflows. Same structural format as the clipboard-mode context file but oriented toward shell usage.

Shell-instructions file entries (shell mode)

Rules for shell-command output:

- Prefer `rg` (ripgrep) over `grep` for content searches
- Prefer `fd` over `find` when available
- Quote paths that may contain spaces using double quotes

Creating a missing file

If the target file does not yet exist:

1. Pick a conventional filename relative to the project root (or ~/.voice/ for global): voice-instructions.md, voice-shell-context.md, voice-shell-vocabulary.md, voice-shell-instructions.md
2. Create the file with a minimal scaffold (top-level heading appropriate to the file's role, then the first entry)
3. Register the new file path under the corresponding key in the relevant voice.json — the project's for project-level, ~/.voice/voice.json for global (create the voice.json if it does not exist)

Guidelines

• Keep each context / instructions file concise — every relevant file is injected into the prompt for its pipeline. For project-level files, aim for under ~80 lines; if a file is approaching that, suggest removing less relevant entries before adding new ones
• Match the formatting style of existing entries, with exact capitalization for each term
• For acronyms, include the expansion in the context file
• Omit widely known terms (JavaScript, React, Git) unless they are specifically being misheard

Step 4: Confirm

After editing, display:

• Whether the project-level or global-level files were updated (and their paths)
• Which mode the change targets (clipboard or shell)
• Which file(s) were modified (vocabulary, context, instructions, or their shell-mode equivalents)
• The specific entries that were added
• If a voice.json (~/.voice/voice.json for global) was updated to register a new file, mention that
• A brief reminder: changes take effect on the next Shama recording (Shama captures file contents per recording — no app restart needed)