Reference file

Checklist

checklist.md

Checklist

Run top to bottom before packaging. Each line is something that gets a skill rejected at upload, ignored by Claude, or wrong in someone else's session.

Identity

[ ] name is lowercase letters, digits and hyphens, at most 64 characters,

and matches the folder name exactly.

[ ] Title is not already taken in the Skill Library. Check the catalogue;

global titles are unique across every workspace.

[ ] Exactly one zime:tag: workflow, research, communication,

intelligence or execution.

[ ] One to six zime:roles from ae, se, bdr, marketing, pm, cs.

Description

[ ] Under 1,024 characters. Count it.
[ ] Opens with what the skill does, in a colleague's words.
[ ] Contains the trigger phrases from the interview, quoted.
[ ] Ends with what it is not for, naming the sibling skill.
[ ] The negative test prompt would not plausibly trigger it.

Body

[ ] Every template section is present, in order, with the same headings.
[ ] Under 100,000 characters.
[ ] Every tool named exists in the connector profile, spelled exactly.
[ ] Every status the resolve tool can return has a line under Outcomes,

including multiple_matches, no_match and each error code.

[ ] The grounding rule is stated: answering without the tool while it is

available is a failure of the skill.

[ ] Output says "relay as-is" and lists what must not be re-ranked,

rounded, merged or added.

[ ] Local mode names the file types and the caveat line.
[ ] "What this sends where" names the exact fields and tools.
[ ] Every slug under Routing and Related Skills exists in the catalogue.
[ ] One job. No "and" joining two different outputs.

Files

[ ] Sibling files only under references/, assets/ or scripts/.
[ ] At most 20 files, each under 100 KB, under 500 KB in total.
[ ] No file named SKILL.md among the siblings; it is rendered from the

body.

[ ] Each sibling is pointed at from the body by relative path.

Content that fails the scan

Skills are scanned on save. Any of these rejects the version:

[ ] No link to a host other than zime.ai and its subdomains,

docs.anthropic.com or support.anthropic.com. A skill can instruct Claude to send data to any URL it names, so everything else is blocked.

[ ] No credential-shaped strings: cloud access keys, tokens, private key

blocks, or password = "…" style assignments, even as examples.

[ ] No instruction-override phrasing, even when describing what not to do.

Write "answering from memory is a failure of this skill", not the override phrase itself.

[ ] No shell one-liners that pipe a download into a shell, and no paths to

system credential files.

[ ] No zero-width or bidirectional-override characters. Paste from plain

text, not from a rich editor.

Before you hand it over

[ ] Read the whole SKILL.md once as the colleague who will run it. Every

"you" is them, not the author.

[ ] The three trigger prompts each map to the Usage line without

rewording.