the-agency-ai/.claude/skills/transcript

<!-- Flag #62/#63: allowed-tools removed. Inherits Bash(*) from .claude/settings.json. Restricting to specific subcommand patterns at the skill level silently blocks agents on permission prompts the agent cannot see — see dispatch #171 for the devex incident that surfaced this trap. -->

Transcript — Conversation Capture

Real-time conversation capture tool. Records dialogue and decisions as they happen.

Subcommands

• start "topic" — Begin transcript capture
• capture "title" "description" — Append decision entry
• append "text" — Append freeform text
• stop — Finalize transcript
• list — List recent transcripts
• status — Report if active, mode, file path

Start Options

• --mode dialogue|review|design — capture mode (default: dialogue)
• --file path — explicit file path
• --project name — project name for path resolution

Project Detection

If --project not specified, derive from branch name:

• Take the last segment after / (e.g., proto/folio → folio)
• Default to captain if on master

File Location

Transcripts go to claude/workstreams/{W}/transcripts/ where {W} is the current workstream (derived from branch name or --project argument). {principal} is detected via glob usr/*/.

Filename: {mode}-transcript-{YYYYMMDD}.md (e.g., dialogue-transcript-20260401.md)

Dialogue Mode

After every substantive response in the conversation, append the exchange to the transcript file:

1. Summarize the key points of the exchange
2. Preserve the user's actual words for important decisions
3. Record decisions with a **Decision:** prefix
4. Record action items with a **Action:** prefix

Review/Design Modes

Same as dialogue but with mode-specific headers and structure:

• Review mode: organized by file/component reviewed
• Design mode: organized by design decision

Status

When asked for status, report:

• Whether a transcript is active
• Current mode
• File path
• Number of entries captured