Here's the training plan laid out: what each session covers, what people should walk away able to do, and the staff survey finding behind each choice. Every session is built around doing, with Claude and Gemini open on everyone's screen, so more than half of each session is hands-on. Training starts once the policy is signed off and accounts are live, so people meet the policy and the tools together.
Six steps, each building on the last. Foundations is shared by everyone; the clinics go deeper by function; the Guides session and the endline set the team up to carry this on.
Groups follow Katrina's suggestion: Group 1 is US-based staff, Group 2 is everyone else (East Africa and South Asia). The clinics are open to drop-ins from any function, and Kim, Dana and Emily joining a couple would be a lovely signal to the team.
Thirteen of you completed the August diagnostic. Four findings shaped this plan more than anything else.
Two more threads run through every session. The team already agrees, 11 to 2, that a person reads the material before AI touches it, so we practise that habit and treat it as settled. And 8 of 13 weren't sure how they'd notice their own thinking drifting toward generic, so we give people a couple of simple, concrete checks for it.
Safe-to-share confidence had the widest spread in the survey, and people's instincts were good while the rules were unwritten. Now the rules exist, the fastest way to make them stick is to use them on real examples together.
| 0:00 | Open. I'll announce the recording and offer the opt-out, which is the policy's own notetaker rule in action. Quick anonymous poll: what did you last use AI for at work? |
| 0:10 | How these tools work, in plain language. A live demo where both tools answer a question about Echidna's public work with total confidence and get something wrong. Why checking matters, and what these tools are good at. |
| 0:25 | The sorting game.HANDS-ON In pairs, sort a deck of around 16 Echidna-shaped examples into green, yellow and red, and name the tool each can go into. Think a grantee's published annual report, a 990, unpublished financials inside a grant report, a draft board deck, a calendar invite with an HR topic in the title, purchased Candid data, a site-visit photo with children in it, your own draft that quotes a grantee. Then we check against the policy and spend our time on the cards where the room split. |
| 0:55 | Which tool, where.HANDS-ON Everyone runs the same ask on the same public document in both tools and compares. Then we practise pointing Gemini at your own Drive versus the web, and phrasing the question for each. |
| 1:15 | Everyday hygiene. What a calendar invite title gives away, how notetakers work under the policy (including pausing the recorder when something sensitive comes up), and a short checklist for moving off personal accounts. |
| 1:25 | Exit ticket. Three quick questions, which also tell me where Foundations II should lean. |
Accuracy was the one thing all thirteen of you named, and over-reliance was the most common worry across both survey waves. This session turns both into habits people practise together, on real material.
| 0:00 | What you tried. A few quick share-outs from the week, wins and flops both. |
| 0:10 | Read-first relay.HANDS-ON In pairs, using a public grantee report: one person reads it and writes their three key points, the other works only from an AI summary. Then compare what each of you caught and what the summary smoothed over. |
| 0:35 | The prompt ladder.HANDS-ON Start from a weak prompt and climb three rungs: add context, name the audience, show the format you want. Watch the output change at each step. |
| 0:55 | Catch the model.HANDS-ON Small teams get AI outputs with errors planted in them (an invented source, a figure that's off, a quote nobody said) and race to find them against the originals. |
| 1:15 | Explain it back, and say so. Can you explain the conclusion without the AI output open? Plus simple wording for noting where AI helped. |
| 1:25 | Commit. Everyone names one real task to try this week and bring to office hours or their clinic. |
Clarity on which tasks AI could help with ranged from 2 to 8, and it tracked function closely. The staff diagnostic called a task-mapping workshop the single highest-value hour of training we could run, so each clinic is one.
These are the tasks each function named in the survey, so they're where we start. People bring their own on the day.
| 0:00 | Map your week.HANDS-ON Everyone lists their recurring tasks and places them on a shared grid: how much could AI help, and which data zone is the material in. |
| 0:15 | Pick two. Each person picks two tasks that are high on help and green or yellow on data. |
| 0:20 | Build it live.HANDS-ON Build a saved prompt, Gem or Claude project for each, test it on real material, and refine it. I'll float between people; Programs, as the largest group, works in pairs. |
| 0:55 | Show and keep. A few people demo what they built, and every prompt goes into the library with a line on when to use it and which zone it's safe for. |
Someone leaves a clinic with a question, they bring it here the same week. The best use is a real task you're stuck on. After the handover, the AI Guides host these.
We'll cover how to run office hours, keep the prompt library useful, handle requests for new tools (the bar Emily set: a meaningful benefit that makes the juice worth the squeeze), keep the policy on its review cadence, and spot drift and raise it kindly.
The same questions as August, so we can see how confidence, clarity and practice have moved. Protected calendar time makes a big difference to response rates.
The session content will get its final examples once the policy is signed off, so the sorting game matches the policy word for word. Please shout if anything here looks off or is missing.
Draft prepared by SA Advisory, October 2026, drawing on the August staff diagnostic, the July and September working sessions, and Katrina's scheduling notes. Comments → aggarwal.shabnam@gmail.com