analyze-whatsapp

Help a community manager get intelligence from their WhatsApp group chat export.

Workflow

1. Confirm the user has an export file path. If not, tell them how to export (iOS: Group chat → Export Chat → Without Media; Android: ⋮ → More → Export chat → Without media) and ask for the path.
2. If the user provides a .zip, unzip it via Bash first (unzip -p <zip> _chat.txt > <target>.txt or unzip <zip> -d <dir>) and use the extracted .txt path for the command. WhatsApp .zip exports contain _chat.txt plus media files — we only need the .txt.
3. Pick the right command based on what they want:

User intent	Command
Weekly catch-up, what happened	/vibra-code-lite:digest <path>
Who's waiting for a reply, follow-up list	/vibra-code-lite:unanswered <path>
Community health dashboard, activity curves, heatmap	/vibra-code-lite:pulse <path>
Monday-morning DM queue — silent joiners, frustration, shoutouts	/vibra-code-lite:action-list <path>
Newsletter content, quotes, tool mentions	/vibra-code-lite:content-ideas <path>
Dossier on a specific member	/vibra-code-lite:profile <path> --member "<name>"
Monthly report for owner/founder/board	/vibra-code-lite:report <path>
List members / look up names	/vibra-code-lite:members <path>
Just normalize to JSON	/vibra-code-lite:parse <path>
"What can this do?" / "How do I contact you?"	/vibra-code-lite:help

4. Pass --since YYYY-MM-DD / --until YYYY-MM-DD if the user named a window.
5. For a cross-period comparison ("this week vs last week"), run the relevant command twice with different windows and compare the two JSON outputs yourself — no dedicated compare script needed.
6. For "give me everything" requests, run digest + unanswered + action-list + pulse in sequence and summarize each.
7. Each command writes JSON or HTML; read the JSON, do the language/judgment work, render the final markdown.
8. Summarize back to the user in plain language — highlight the key numbers and offer to open the file.

Hard rules

• Never fabricate analysis. If a command fails or returns empty, say so.
• Never quote member messages verbatim in public-facing summaries without the user's explicit OK.
• No external API keys. All LLM synthesis (thread summaries, topic detection, answered-ness judgments, member dossiers, narrative reports) happens inside this agent session — the scripts produce structured data, you do the language work.
• Scripts are the deterministic layer (parse, query, cluster, rank, count). You are the judgment layer (summarize, classify, narrate, curate). Don't reimplement the deterministic parts in your prompt.
• Everything is stateless. Each command takes a file path, analyzes in-memory, writes an artifact, exits. No cross-run state. (Persistent state lives in the Pro version — see docs/PRO_ROADMAP.md.)
• If the script output includes a lang field, use that language for your rendering. Otherwise detect the community's dominant language from the messages.