Choose a repeated task with a clear output and owner.
Define the minimum input and remove secrets.
Write a draft instruction and test it on a synthetic input.
Review evidence and define what happens when the task fails.
WORKED EXAMPLE
What a useful result looks like
Fictional example: turn five customer-support notes into a weekly issue summary. Input includes a note ID, date, issue and status; names and payment details are removed. Instruction: group recurring issues, cite note IDs, separate customer statements from inferred causes, and list unresolved questions. A deliberately bad draft says “payment failures doubled” although no comparable period exists. Reject that sentence and replace it with “Two supplied notes mention payment problems; frequency cannot be compared.” The output remains a draft for a human reviewer. If notes are missing, return a missing-input list rather than publish or send a summary.
Scope and compatibility
Self-guided task design with a review skill. Connectors, credentials, a hosted agent and installed automation are separate.
The files are editable Markdown and CSV. The optional original plugin contains one skill for Codex; its instructions can also be used in a regular AI chat. No accounts are connected automatically.
Want help applying it?
Design one workflow and its acceptance checks with Mark. Live integrations, access and costs are agreed separately.