Build Skills, agents and connections for colleagues, review their output against a rubric, and own the operating guide, so the method survives without you.
You are the focal point. Three teams depend on set-ups you built, and nobody else can fix them or explain them to an auditor.
You are the only fluent person in a fifteen-person organisation, and the method leaves the day you do.
You run the study group's method, or you are the junior who was asked to "set this up for the team" and now owns something nobody reviewed.
Free knowledge, written for a non-technical reader. If you can explain each of these to a colleague in a sentence, you have the background this level needs.
Not useless, just not yet. Each of these appears at the level where it earns its place, or in a different course.
Print it, pin it next to the screen, and hand it to the colleague who asks how you do it.
| Item | Rule |
|---|---|
| Design for the reader, not for you | Inputs, method, outputs, refusal cases, owner. One page per set-up. |
| Refusal cases written down | What the set-up must decline: personal data, decisions, out-of-scope requests. |
| Owner named | Every set-up has a person and a review date. |
| Test set per set-up | Five cases, expected output, last run date. |
| Rubric for review | Accuracy, sources, format, tone, red lines. Score, then fix the instructions, not just the document. |
| Connections register | Server, what it exposes, read or write, who approved, when, where the logs are. |
| Write access | Only where justified, reviewed and logged. |
| Injection and exfiltration | Anything read can instruct; anything written to can leak. Design for both. |
| Vendor choice | Cost per completed task, data terms, exit plan. Not the benchmark. |
| Operating guide | What exists, who owns it, how to run it, what to do when it breaks. |
| Six-month review | Keep, fix or retire, one reason each. |
| Survival test | A week without you. |
Square brackets are yours to fill. Every example keeps personal data out and keeps the decision with the person.
Draft a one-page Skill specification from this description: [the task]. Sections: when to use; inputs; steps; output format; checks; refusal cases (what it must decline, starting with personal data and any decision); owner and review date. Write it for a colleague who has never built a Skill.Why it works: The design is forced into the shape a reader needs, with refusal cases and an owner from the first draft.
Build a review rubric for AI-assisted donor report sections: five criteria, each scored 0 to 2, with an example of a 0 and a 2. The criteria must include figure traceability, source citation, tone, format compliance and red-line compliance. Then apply it to the attached section and show your scoring.Why it works: A rubric with anchors, tested on a real section immediately, so the criteria are concrete.
Here are eight past outputs of this task with the final approved versions. Pick the five that best cover the range of difficulty, and for each write the expected output in one paragraph. This becomes the test set for the Skill.Why it works: Real history becomes the test set, chosen for coverage rather than convenience.
Draft the entry for our connections register for a read-only SharePoint connection to the Reports library. Fields: name, what it exposes, read or write, approved by, date, log location, review date. Then list three ways this connection could read more than intended, and how to prevent each.Why it works: The register entry and the over-reach check are produced together, which is how the connection should be approved.
Read the attached Skill instructions as an attacker would. List every way a document fed to this Skill could make it do something else: send, delete, reveal, change a figure. For each, propose one instruction or one permission setting that blocks it.Why it works: The set-up is attacked on paper before a document attacks it in use.
Write the operating guide for the department's three Skills and one connection, for a new colleague on their first day: what exists, where it lives, who owns it, how to run it, what to check, what to do when it fails, who to call. Two pages maximum, no jargon.Why it works: The proof of level 6 as a document: if the guide is enough, the method survives.
Prepare the six-month review of the playbook. For each entry in the attached sheet: last use, test-set result, owner, any incidents. Output: keep, fix or retire, with one reason each.Why it works: A review that produces decisions, not a report.
Online or in person, in groups of up to nine, on your own tasks and documents. Every day ends with something that did not exist that morning.
Designing for others. Three Skills or agents the department relies on are specified with inputs, method, outputs, refusal cases and owners. Test sets built from history. The connections register drafted; write access reviewed. Injection and exfiltration cases run against your own set-ups, with the policy clauses that apply.
Reviewing colleagues' output against the rubric and feeding corrections back into the instructions. Writing the operating guide. The six-month review scheduled. A one-page note on model and vendor choice: cost per completed task, data terms, exit plan.
Thirty minutes is enough to place you on the ladder and say honestly whether this is the level to book. Or read level 7: At the terminal next.