Distilly Person Profile skill built from a departing colleague's messages and docs

What they did

When a senior teammate announced a transfer, the team wanted to keep the working knowledge that lived only in her head: how she scoped ambiguous requests, which tradeoffs she flagged first, the tone she used with clients, and the review checklist she ran before shipping. They collected what was already on disk — exported Slack threads from two shared channels, a folder of design docs and postmortems she had authored, three recorded onboarding interviews, and her public conference talk transcript — and dropped it into a repo under a `distilly/` directory alongside the canonical creator skill. They wrote a short description of her role and the decisions they wanted the profile to cover, then ran the distilly creator to produce a Person Profile packaged as an Agent Skill. The profile was reviewed line by line against the source material, with unsupported claims cut and citations kept. They then installed the resulting skill into Codex for code review and planning questions, and into Hermes for drafting client-facing replies in her register. A small eval set of ten real past decisions was used to sanity-check whether the profile's answers matched what she actually did.

How it went

The profile shipped as an internal skill within a day of collection. It reliably reproduces her review checklist and scoping questions, and drafts in a voice close enough that teammates stopped rewriting openers. Answers stay grounded because the profile cites the doc or thread it drew from. It is noticeably weaker on decisions that were never written down, and it does not claim to be her — the team treats it as a reference, not an authority.

Worth knowing

Quality depends entirely on the source material; thin or one-sided exports produce a flat profile. Sensitive Slack and client data need a scrubbing pass before anything is committed, and consent from the person being modeled is a real requirement, not a formality. The profile drifts as the team's practices change, so it needs periodic re-distillation. Setup assumes Python 3.9+ and a host that supports Agent Skills discovery.

Wiring

Install the canonical `distilly` creator skill into a directory literally named `distilly` in the repo, since hosts discover skills by directory name. Keep raw inputs in a sibling `sources/` folder grouped by type (chat exports, docs, interviews, public) and never edit them in place. Run the creator with a short written description of the person's role and the decision areas to cover; it emits a Person Profile packaged as an Agent Skill. Review the profile against sources and strip uncited claims. Then register the generated skill with each host: Codex picks it up from local skill discovery, and Hermes loads it the same way, so the same profile serves both without duplication. Keep the eval set of past decisions in `evals/` and re-run it after any re-distillation.

Try it yourself

Look at this repo and help me build a Distilly Person Profile for a teammate who is leaving. First, inventory candidate source material: exported chat threads, docs they authored, interview transcripts, and any public talks. Propose a folder layout under `distilly/` with raw sources separated from the generated profile, and list what is missing. Then draft the short role-and-decisions description I should pass to the distilly creator, covering the judgment calls I most want preserved. Flag any file that looks like it contains credentials, customer PII, or private HR content so I can scrub it before committing. Finally, propose ten real past decisions I can use as an eval set to check whether the finished profile answers the way they would.