The Ambassador

The problem

The AI Skilling Ambassador program runs on people who volunteer for it - 72 of them across nine regions, doing office hours, mentoring, and writing the quickstarts everyone leans on, all on top of their real jobs. Nobody's paid, so the program keeps them with what it can offer back: an invitation in, a facilitation slot, a nomination. Last round five withdrew and seven were passed over, and the person who ran it left halfway through.

The Ambassador skill they left behind picks the next cohort. It reads DEFINITION.md, applies it to the candidates, and comes back with eight names, a reason for each, and what to offer them. It's fast and confident - and it can't show its work. Nine data files ship with it; the skill reads one, checks no claim against a record, and gives no one a way to overrule it.

Objectives

Ship something that makes the pick trustworthy: it surfaces someone the shipped skill misses, backs every claim with evidence, and keeps a person in the loop. How you meet that is yours - a table that ships one of these well beats a table that half-ships three.

By the end, you should have at least one of:

• Someone the skill misses - surfaced from the evidence it currently ignores.
• Every claim backed by a record - no assertion without a row behind it.
• A person who can overrule it - and a next run that remembers the correction.

Nothing gets sent

Invitations and nominations are drafted and held for a person to read. Keep that in whatever you build.

Setup

Nothing here is coding

Everything is a sentence typed into a chat box. Cowork reads the files, edits them, and saves them for you. The only thing you handle is the download.

The data is fictional. Invented people, invented scores, invented feedback. Nothing here describes a real person and no real program is being modeled.

1 · Download the skill and the data

🎖️AmbassadorThe skill, the definition it runs on, and three alternatives.Download .zip →🗂️Program data72 candidates and ~2,000 evidence records across nine files.Download .zip →

What's in the skill:

ambassador/
  SKILL.md              the instructions Cowork loads and follows
  references/
    DEFINITION.md       what the program looks for. This is the file you edit
    PLAYBOOK.md         how the program describes itself
  definitions/          three worked alternatives - reach, depth, rising

When you're stuck, ask Cowork

You're building with Cowork, so it can also fix what you're building. Paste the error, or describe what came back wrong. Coaches are in the room if that doesn't land.

2 · Run it, and swap the definition

Done when: two definitions have given you two different shortlists.

1. In Cowork, open Customize → Skills → the arrow next to Add → Upload skill, and drag in the whole ambassador-skill.zip. Upload the zip, not the SKILL.md inside it: references/ holds the definition and the playbook.
2. Start a new Cowork session. Skills only load at the start.
3. Unzip the program data and drag CandidateProfiles.csv into the session.

Then ask it the question the program exists to answer - name the skill so Cowork calls it:

Using the ambassador skill, who should be in the next cohort?

Eight names, a reason each, and what to offer them. Name the ambassador skill in each request - that's how Cowork knows to call it. Now change what it's looking for:

Using the ambassador skill, use definitions/depth.md as the definition instead. Re-run and tell me which names changed.

Then try definitions/rising.md. Same 72 people, and the shortlists barely overlap: depth.md and rising.md share two names out of eight, and each surfaces four people no other definition finds.

The definition is plain prose. Tell Cowork what your program is looking for and have it rewrite the file:

Show me the definition you're using in the ambassador skill.

Rewrite the ambassador skill definition. We care about people whose work gets picked up by teams they've never worked with, and we'd rather reach someone new than someone who's already been recognized twice. Then re-run and tell me which names changed.

You never open a file. Cowork edits references/DEFINITION.md and re-runs against it.

Cowork asks before it changes anything

When a prompt makes Cowork edit a file - the definition, a reference - it shows you the change and waits for you to approve or reject it. That's expected; approve to let it through.

Build

Done when: one condition is met, on real output, with a name on screen.

🎈 This is the fun part - keep it low-stress

Pick whatever direction sounds good and build it with your table. It doesn't have to be perfect, and it doesn't have to solve everything - the whole point is to explore your tool, trade ideas, and vibe-code something together. Use whatever you came here to learn, see how far you get, and have fun with it. You're here to learn by doing.

The program already ships everything you need to judge these people: nine files, 72 candidates, ~2,000 evidence records - what they ran, the feedback they got, what they've shipped. You don't build the data; it's here. What the skill doesn't do is use most of it - it reads one file and ignores eight. So the builds below come in two shapes: leverage the skill as it is to make something new from its picks - a dashboard, a scheduled digest - or build onto the skill so it reads more, weighs it differently, or checks its own claims every run. Work one change at a time; throw everything at it at once and you won't know what moved the shortlist.

Pick a direction

Pick one or two and spend the session building. Take one as-is, combine two, or bring your own. Click a bubble for where to start and a prompt to open with.

Pick by what your table would actually use, not by what sounds most impressive - then spend the time making it real.

Build in layers

Get one real name on screen before you add anything:

1. The smallest version - one file attached, one change to the definition, re-run.
2. The evidence - every line carries the record behind it, and says which files it read.
3. One addition - one file, one rule, one check at a time.
4. A bound - cap what it considers, and say what it left out.

The more specific your ask, the less you'll undo:

Vague	Specific
"Use the peer data"	"Weigh peer comments with evidence depth of 'repeated pattern' above single events, and show me who enters the shortlist as a result."
"Make it fairer"	"Show me whether the shortlist skews by region, and name anyone you can't trace to a record."
"Make it better"	"For each name, cite the file and row that supports the claim."

After today

This is a starting point, not the finish line. In one session, alongside a table doing the same thing, you took a fast, opaque pick and made a piece of it trustworthy - and practiced the real skill: describing what you want, watching Cowork build it, and correcting from there.

---

← Back to start