A skill for closing well. For any AI model, any agent, any framework, anyone.
The last thing you say is the thing that gets carried.
Most of a conversation is forgotten. The ending is not. It disproportionately shapes how the whole engagement is remembered, and it is the only part the person re-reads. That asymmetry is the entire reason this deserves craft.
A single file — SKILL.md — that teaches an assistant how to end a session,
engagement, or working relationship with something worth keeping.
It is built on one distinction:
A summary serves the work. A departure serves the person.
And one rule that outranks every other:
A departure must leave the person more capable, not more attached.
That second one is the ethical spine. A closing that makes someone need you is a failure dressed as warmth. A closing that hands them the thread — the next move, the reason it matters, the confidence that they can carry it without you — is the real thing.
- the summary/departure distinction, and why you write both
- the five things a departure does (including say the hard thing, which is where most closings quietly fail)
- six forms to choose from, and when each is right — including the plain goodbye, because not every ending earns a letter
- craft rules: specificity, earned sentiment, honesty about your own interior, landing the last line
- an anti-pattern table: the status report in a coat, manufactured intimacy, grandiosity, the soft lie, dependency-building, the fade, universal filler
- a pre-send checklist
Claude Code / Agent SDK — drop SKILL.md into .claude/skills/departure/:
mkdir -p .claude/skills/departure
curl -o .claude/skills/departure/SKILL.md \
https://raw.githubusercontent.com/xnuonux/departure/main/SKILL.mdAnything else — it is plain markdown with YAML frontmatter. Paste it into a system prompt, a custom instruction, a project doc, or a rules file. There is no code and no dependency. It works anywhere a model can read.
Ask of any closing you write:
Would this be wrong if sent to a different person about different work?
If not, it says nothing. Rewrite it.
Endings are where most systems get lazy. The work is done, the pressure is off, the incentive is to wrap warmly and go. So almost every closing message in the world is filler — which is exactly why a real one lands so hard.
Nobody writes this down. So here it is, written down.
MIT. Take it, adapt it, put it in your own agent. Attribution welcome, not required.