Inspiration
In Sweden, we have the concept of döstädning (death cleaning): sorting through your belongings while you are still here so your family does not have to guess what matters when you are gone. Physical homes get cleaned out, but local NAS drives quietly fill up with terabytes of unsorted files.
The hardest part of a digital handover is not locating files; it is transferring the context that only lives in the owner's head. A drive can hold tens of thousands of photos and audio clips, yet only the owner knows who the unnamed faces are, which voice recording has a parent's voice, which contracts are still active, and which folders can be thrown away. Without a guide, families are afraid to delete anything. I built When I am Gone to let a NAS owner pass down that knowledge alongside an archive their heirs can verify.
What I learned
Designing for an end-of-life handover requires very different rules from everyday file automation:
- Explicit approval over automation: In a handover archive, a made-up sentence or a wrong assumption is worse than a missing file.
- Two personas separated by time: A handover skill serves two personas with different needs. Prepare structured questions to extract knowledge from the owner. Verify giving heirs a clear answer on whether anything has changed.
- Offline verification: Cloud estate services require ongoing subscriptions and third-party trust. Writing verification in physical form together with a personal "letter from the other side" counters that.
How I built it
The skill is a pure orchestration pipeline across four UGOS experts (file-agent, document-agent, photo-agent, and knowledge-agent) using 23 delegations:
- Structure and modules: I split the skill into Prepare and Verify flows.
- Non-destructive by design: A handover tool cannot risk deleting memories. The skill contains zero
deletecalls. - Contract tests and simulation: To verify the logic before running on live UGOS agents, I wrote automated contract tests an walkthrough files covering simulated cases, plus a synthetic family dataset.
Challenges I faced
- State handoffs across 23 delegations: Chaining four experts means an empty or partial result in step 3 can break step 15. Every step is built to use only the real output from the step before it, and all five states are handled explicitly. If an optional module fails or returns nothing, the rest of the archive still completes.
- Unknown expert payloads: Building against the kit without live NAS hardware meant some expert behaviours could not be observed directly: how
scan-dedupegroups duplicates, howfigure-albumsreturns unnamed face clusters, and howsave-textnames files. I handled those unknowns with defensive fallbacks. - Keeping the letter factual: Language models naturally try to add emotional filler when drafting personal letters. Constraining
document-agentin the writing step to stick to approved registry facts and quote the owner's words unchanged took strict instruction boundaries and walkthrough testing.
