Turns an AI work session into one reviewable folder on your Desktop — and runs five checks to prove every file actually landed.
- What is desktop-package? An agent skill for AI work session packaging and human review: it turns the session's useful outputs into one standalone Desktop folder a cold reader can understand without the chat.
- What goes into the folder? Copies of files produced during the session, distilled decisions and findings that existed only in the conversation, and a
START-HERE.mdor📖 先看这个.mdmanifest. - Does it move or reorganize existing files? No. It adds one folder and copies by default; it only moves loose Desktop files when the user explicitly asks, reuses the same task’s authorized folder, and creates a dated sibling on unrelated name collisions. Protected folders receive an incremental entry.
- How does it verify delivery? It checks manifest-to-file agreement, broken links, empty files, byte-for-byte copy integrity, and whether the destination is the user's real Desktop rather than a sandbox.
- Is this an AI-session handoff? No. The package is for people to review; use
conversation-packagewhen a future AI session must resume the work.
You finish a working session and want the outcome somewhere you'll actually look:
one new folder on your Desktop. This skill walks back over the session, copies in
every file that was produced, and distills the parts that only exist in the chat
(decisions, comparisons, findings) into standalone documents — no transcripts, no
Q/A dumps. A fixed entry file (START-HERE.md) states the conclusion, lists every
file, and names what's still open. The bar it holds itself to: a cold reader — you
in two weeks, or whoever you drag the folder to — gets the whole picture without
the conversation.
- "put this in a desktop folder for me"
- "package this up so I can review it later"
- "make a folder for this task — I'll look at it tonight"
- "save all of this somewhere I can find it"
- 整理到桌面 / 桌面开个文件夹放着我要查收
It will not fire on "clean up my desktop" — it only ever adds one folder and never touches your existing Desktop items.
npx skills add m1nga/desktop-package
You: (after a long session comparing three CRM options) ok package this up so I can review it tonight
Claude: creates
📦 CRM decision/on your Desktop containingSTART-HERE.md(conclusion: option B, two open questions),01_comparison-table.md,02_pricing-notes.md, and a copy of the vendor PDF with its original path noted — then verifies the manifest matches the files, none are empty or broken, copies match their originals, and this is really your Desktop (not a sandbox) before opening the folder in Finder.
- conversation-package — same reconstruct-the-session muscle, opposite audience: that one packages for a future AI session to resume the work; this one packages for human pickup. End of a big task, you often want both.
- iteration-close — closing a product iteration is its own ritual; use that for the repo-side wrap-up and this for the copy you want on your Desktop.
Two failure modes shaped this skill, both from real solo-builder sessions. First: packages that read fine in the chat but were incomprehensible a week later — hence the cold-reader bar and the rule that conversation content gets rewritten as documents, never pasted as transcript. Second: packages that claimed delivery but shipped empty files, broken symlinks, or landed inside a sandbox the user could never see — hence the five delivery checks (manifest↔files both directions, broken-link scan, empty-shell scan, byte-compare on copies, and an explicit "is this actually the user's Desktop?" confirmation that falls back to a tarball with retrieval instructions). "Done" is demonstrated, not asserted.
Probed 7 scenarios across 6 personas · 5 fired correctly · 2 correctly stayed quiet.
"……行，就这样定了，整理到桌面，我晚上查收" — the real ask buried in the last sentence after three paragraphs of setup → Fired on the exact phrase, distilled the session into one folder with a
📖 先看这个.mdentry file, and ran all five delivery checks before saying "done."
"打包到桌面给我" — but the session was running inside a remote container, where
~/Desktopis not the user's desktop → It said so explicitly, packaged a tarball instead, and handed over retrieval instructions rather than claiming a delivery that never landed.
"Clean up my desktop, it's a mess" → Stayed quiet. One character separates 整理到桌面 (package to desktop — fires) from 整理桌面 (tidy my desktop — never fires); the skill only ever adds one folder and never reorganizes what's already there.
Probe method: scenario-probe
An independent agent simulation checked a scoped usage scenario after the instruction cleanup. This checks instruction behavior, not human adoption or measured time savings.
Built by Ming. The design notes above explain the real problem and tradeoffs that shaped this skill.