build: Skill for Claude Code

Process Document

Convert a raw document into a structured, searchable markdown summary.

Usage

/process-document projects/{project}/raw/{filename}

Transcript detection

Detect call transcripts via two signals:

1. Path: folder name contains call-transcripts OR filename contains recording, transcript, sync, meeting, call
2. Content: 3+ unique speaker labels in first 50 lines, in either format:
   • **Speaker Name** (newer .md transcripts)
   • Speaker Name 0:00 (older .txt with timestamps)

Either signal is sufficient. Both together = high confidence.

Workflow

1. Read the source document
2. If transcript detected: run the transcript cleaning sub-step (see below), then use the cleaned file as input for step 3
3. Generate a structured markdown summary (see format below)
4. Save to projects/{project}/docs/ with a matching name (e.g., raw/spec.pdf becomes docs/spec.md)
5. Update projects/{project}/docs/INDEX.md with a one-line description
6. If transcript was cleaned: include both raw and cleaned source paths in the summary header

Transcript cleaning

When a transcript is detected, before summarizing:

1. Read transcript-cleaning.md (relative to this skill) for the full cleaning instructions
2. Use the Agent tool to spawn a cleaning sub-agent:
   • model: "sonnet"
   • subagent_type: "general-purpose"
   • Prompt must include:
     • The cleaning instructions from the reference file
     • The full raw transcript content
     • Paths to project context files for the sub-agent to read: projects/{project}/docs/transcript-context.md, projects/{project}/docs/glossary.md, projects/{project}/BRIEF.md
     • The original filename (for the source frontmatter field)
   • The sub-agent reads context files, applies ASR corrections, and returns the cleaned transcript
   • For non-English transcripts, the sub-agent also translates to English (per the cleaning instructions)
3. Save the sub-agent's output as {original_path}/{original_name}_cleaned.md (same folder as the raw file)
4. Proceed with the cleaned transcript as input to the summary step