Use immediately when the user asks you to do something net-new -- a task you haven't done before, no existing skill applies, and getting it right will require nontrivial research, exploration, or experimentation. When doing this, give a very short confirmation message to the user's request, then load this immediately before responding further. Your confirmation message shouldn't mention loading the skill. Skip when an applicable single skill already exists or for pure dev/code-writing work.
Use when you want to create a new web view for the user. Covers scaffolding a new FastAPI service (canonical path) and the escape hatch for wrapping a pre-existing third-party server, plus diagnostic references when things misbehave.
Create a sub-agent to perform a larger task. Use when work is large enough to warrant a separate context, involves multi-file changes, or benefits from isolation.
Use whenever you want to use latchkey commands or interact with third-party or self-hosted services (Slack, Google Workspace, Dropbox, GitHub, Linear, Coolify...) using their HTTP APIs on the user's behalf.
Push local improvements to shared infrastructure (skills, scripts, CLAUDE.md scaffolding, Dockerfile, services.toml) back to the parent template repo so other agents derived from the template benefit. Opens a separate per-feature PR per logical fix; never pushes directly to upstream `main`. Do not push agent-specific content (PURPOSE.md, memory, runtime state). For pulling updates from upstream, use the `update-self` skill instead.
Extend, refactor, or verify a crystallized skill under `.agents/skills/`, or a shared script or reference under `.agents/shared/` that other skills consume. Invoke at turn-end when you had to do additional repeatable work around the artifact (absorb flow), or when you and the user discussed a change and you applied it live (verify flow).
| name | file-sharing |
| description | Use to read and write files and directories on the user's local filesystem. |
Use this skill when the user asks you to work with files located directly on their PC.
latchkey curl calls to communicate with the remote WebDAV server. (They are the same as normal curl calls, just going through the Latchkey Gateway.)latchkey curl -XPOST http://latchkey-self.invalid/permission-requests when the curl request comes back with the "request not permitted by the user" message. See the "Ask for user permission" example below.The base URL is http://latchkey-self.invalid/minds-api-proxy/api/v1/files. Only the user's home directory and the user's system temp directory are accessible. MOVE and COPY operations are not supported.
latchkey curl -O http://latchkey-self.invalid/minds-api-proxy/api/v1/files/home/hynek/project/notes.txt
latchkey curl -T localfile.txt http://latchkey-self.invalid/minds-api-proxy/api/v1/files/home/hynek/project/remotefile.txt
latchkey curl -s -X PROPFIND -H "Depth: 1" http://latchkey-self.invalid/minds-api-proxy/api/v1/files/home/hynek/project/ | xmlstarlet sel -N d=DAV: -t -m "//d:response/d:href" -v . -n
When a request comes back with the "request not permitted by the user" message, ask the user for permission first:
# 1. Retrieve the list of your existing permissions if necessary.
latchkey curl http://latchkey-self.invalid/permissions/self | jq .rules
# 2. Ask for the necessary missing permissions.
latchkey curl -XPOST http://latchkey-self.invalid/permission-requests \
-H 'Content-Type: application/json' \
-d '{"agent_id": "'"$MNGR_AGENT_ID"'", "type": "file-sharing", "payload": {"path": "/home/hynek/project", "access": "READ"}, "rationale": "I'"'"'d like to access the /home/hynek/project directory in order to find the most recent accounting spreadsheet you asked me about."}'
The body must be a JSON object with exactly four fields:
agent_id (use $MNGR_AGENT_ID), rationale, type (use "file-sharing"), and payload.
payload must be an object with exactly two string fields: path and access. path must be absolute, access must be "READ" or "WRITE".
After posting, wait for a system message indicating whether the user approved or denied the permission request.