AICOURSE.TOP — TEACHER QUICK PACK ================================= Purpose ------- Teach beginners to distinguish decisions fixed in code from decisions handed to a model, and to support that choice with a small evaluation and an explicit control boundary. Before the room --------------- 1. Open /en/handbook/ and /en/lab/ on the classroom browser and network. 2. If using live calls, create a separate low-credit, revocable DeepSeek key. 3. Run Save & test, one first call, the three-order preview, and one full Eval. 4. Compare the model, usage and charge with the Provider dashboard. 5. Prepare the no-key route: scripted Handbook + Lab step 2 + projected results. 6. Never share the teacher key. Forget clears the tab; revocation happens at the Provider. Offline Stage 0 --------------- Open /en/build/ for the complete setup and recovery path. In a local clone: npm ci npm run course:offline Then open course/stage0-hello/run.ts and replace the empty QUESTION string with any non-empty question before running Stage 0: npx tsx course/stage0-hello/run.ts --offline npx tsx course/check.ts 0 --offline The checker succeeds with "stage 0 complete." This path uses the bundled deterministic stand-in: no Provider key, request, private recording or charge. It proves the local control flow, not live-model variation. Call disclosure --------------- Step 1: 1 paid request, max 250 output tokens. Step 3: up to 3 paid requests, max 3 x 300 output tokens. Eval: 20 generator + up to 8 judge = 28 calls, max 7,600 output tokens, concurrency 4. Recommended journey with two Evals: up to 60 calls / 16,350 output tokens. Each additional Eval can add 28 calls. Stop prevents new scheduling and attempts to abort up to four in-flight calls; an in-flight call may still be billed. 45 / 90 / 180 minutes --------------------- 45: rules wall, teacher-led prompt preview, baseline Eval, add context, re-run. 90: add Handbook prompt/context/eval, loop/graph, security and decision tree. 180: learners run the Lab with their own keys or the no-key route, play the Control Room, then begin the local TypeScript course. Projection cues --------------- 1. On the Handbook dial ask, "Who chooses the next step, and when?" 2. At the rules wall ask for one phrasing nobody anticipated; do not explain the wall before learners feel it. 3. Before the first Eval, stop and collect predictions. Let the baseline score land before discussing it. 4. Add only the menu, re-run, and ask which failure changed and which did not. 5. At loop/graph ask which decisions the model may make and which code or a person must retain. 6. At security ask where an untrusted email stops being data and starts trying to issue instructions. Reference-answer boundary ------------------------- There is no single approved Agent design. A defensible answer identifies an observable failure, a proportionate control and the trade-off it creates. `course/SOLUTIONS.md` is one complete implementation, not the only acceptable one. Use the rubric below to assess evidence rather than framework choice. Learner worksheet ----------------- 1. Give one phrasing the current rules miss and the extra rule it would need. 2. Before Eval 1, predict the score and give one reason. 3. Record the baseline, one failed case, and the likely failure source. 4. Name one change and why it should affect that case. 5. Before Eval 2, predict the direction and size of change. 6. Record the new score, one improvement, and one regression. 7. Name one decision for the model and one retained by code or a person. 8. Transfer to a domain you know: failure input, minimal eval, irreversible gate. Project rubric (0 absent, 1 partial, 2 evidenced) ------------------------------------------------- - Problem boundary: goal, non-goals, and model authority. - Failure input: concrete, realistic, reproducible. - Minimal eval: cases, expected evidence, and inspectable scoring. - Irreversible gate: code/person control before the action. - Trust boundaries: Provider, tools, secrets, and untrusted inputs. - Reflection: failures, regressions, cost, trade-offs, and next test. Acceptable evidence includes secret-free console output, small JSON/Markdown results, a code link, one failed-case trace, or a local report card. There is no single approved framework, Provider, or prompt; assess observable evidence and the quality of the trade-off. Privacy ------- Keep observation sheets local. Never put learner keys, prompts, raw Provider bodies, replies, or identifiable work into the site, screenshots, or CI logs. Full facilitation plan: https://github.com/HUDongpin/agent-edu/blob/main/TEACHING.md Offline setup: https://aicourse.top/en/build/ Reference implementation: https://github.com/HUDongpin/agent-edu/blob/main/course/SOLUTIONS.md Licence: MIT.