The Skill and Prompt Registry

Use this when more than one person relies on the same prompts or skills, or when you want your own to survive the next model release. One entry per prompt or skill records who owns it, how mature it is, the standard its output has to meet and a few test items you can rerun. When a new model arrives, the test items tell you in an hour what held and what slipped.

The frame

One entry per prompt or skill.

Name: [what the team calls it]
Task: [the recurring work it does]
Rung: [1 raw prompt / 2 saved prompt or library entry / 3 skill / 4 runs as part of a workflow]
Owner: [who keeps it current]
Version and date: [v1.2, 3 October]
Where it lives: [the shared folder, skills library or tool]
Input it needs: [the sources a user supplies each time]
Standard: [one or two lines, taken from the corrections you usually make]
Test items: [two or three past inputs with the output that met the standard]
Last tested: [date and model]
Result: [Holds / Slipped, with one line on what changed]
Recheck when: [a new model, a format change, or every quarter]

Have an AI turn your saved prompts into entries

Below are prompts and skills my team uses. For each, write a registry entry with these fields: Name, Task, Rung, Owner, Version and date, Where it lives, Input it needs, Standard, Test items, Last tested, Result, Recheck when.

Rung is 1 for a prompt typed fresh each time, 2 for a saved prompt or library entry, 3 for a skill that runs from stored instructions, 4 for one that runs inside a larger workflow.

Write "MISSING:" for any field my notes don't cover. For Standard, suggest two lines based on what the prompt asks for, marked "suggested", for me to confirm against the corrections we make.

[paste the prompts or skills, with any notes on who uses them]

Strip client names and commercially sensitive detail from any test items before they go into the registry or a third-party AI tool.

Worked examples

The entries below are illustrative.

Friday status update skill

Name: Friday status update
Task: Turns the week's workstream notes into the sponsor update
Rung: 3 skill
Owner: PM
Version and date: v2.0, 26 September
Where it lives: Team skills folder
Input it needs: Workstream notes, tracker export, last week's update
Standard: The decision needed is in the first line. Every date that moved names who confirmed it. Under 200 words.
Test items: Weeks 18, 21 and 24, each with the update that went out
Last tested: 6 October, on the newly released model
Result: Holds. One fewer edit: it now puts the decision first unprompted.
Recheck when: Next model release, or if the sponsor changes the format

Risk register synthesis prompt

Name: Risk register synthesis
Task: Summarises changes in the RAID log for the fortnightly risk review
Rung: 2 saved prompt
Owner: Project coordinator
Version and date: v1.1, 12 August
Where it lives: Shared prompt library, risk section
Input it needs: RAID log export with review dates
Standard: Every risk named cites a log ID. New, closed and re-rated risks listed separately.
Test items: The 29 July and 26 August reviews
Last tested: 12 August
Result: Slipped once in testing: a re-rated risk appeared under new. Instruction added.
Recheck when: Quarterly. Candidate for rung 3 once it holds for two cycles.

Notes on the fields

Rung comes from the maturity ladder: prompts, a saved library, skills, then orchestration. The registry shows which tasks are still re-briefed by hand.

Standard is the line between a prompt that looks good and one that is validated. Write it from the edits you make to the output.

Test items turn a model change from a guess into a short test. Rerun them and compare with the outputs that met the standard.

Check an entry before it goes in

• Has the prompt produced an output that met the standard, with minimal editing, on a real recent input?
• Is there a named owner who isn’t the only person who uses it?
• Are the test items stripped of names and sensitive figures?
• Is there a Recheck when trigger?

When it does not apply

A prompt you use once, or only for your own exploration, doesn’t need an entry. Add it when someone else relies on it, or when you’d want to know if a new model broke it.

What this feeds

Each skill that runs on a schedule also gets an entry in the AI Delegation Register, where its owner is the config owner. The test items follow the method in How to Test a New AI Model on Your Own Project Work, and the rungs come from AI Skills for Project Managers.