Intake automation

The AI context system is only as current as what goes into it, and until now everything went in by hand. WhatsApp went in as screenshots, calls as two PDFs each, mail printed to PDF, LinkedIn profiles saved as a PDF of the first page. Over the six weeks before this, that was about sixty chat captures, forty-five call transcripts, thirty-five profiles and twenty mails.

The feeding was the bottleneck, not the thinking. What I had not fed in, Claude could not know.

Eight intake streams, built in one day, each pulling from where the material already lives and filing it as text under the right client or project.

Scheduled capture. WhatsApp chats are read from WhatsApp Web three times a day, into one history per contact. Mail I label, and mail I send to people I work with, comes in once a day. Call transcripts and notes come in as text from the recorder, so "debrief my call with X" fetches the call and runs the debrief in one step.

On request. I send a LinkedIn link and the full profile is filed, every role with its description. "Status on X" returns an overview plus a short version I can send.

The morning brief. At 07:30 the captures run first, then one page tells me what came in, what is waiting on me, what is due this week, and which calendar entries have no matching commitment on file.

Every chat, inbox and recorder has a bookmark, so nothing is pulled twice and nothing is missed. A voice note or video with no transcript is listed for me instead of silently skipped. The runs are Claude skills on a schedule in the Claude app, mostly on Sonnet, and they save their work locally. Nothing is pushed anywhere.

What I chose not to build: automatic reply drafts, which pile up unread, and automatic weekly reports. Both exist on request.

The first scheduled runs went through the same day. The morning brief arrived with four chats, two call recordings and four labelled mails already filed, and it listed a promise I had made in a chat that morning as due that day.

My estimate from those six weeks is about four hours a week saved, which is a count times minutes, not a measurement. The runs only happen while my laptop is awake, and that is the next thing to solve.

What are you still feeding your own AI setup by hand?