TutorialsAssignment: Turn a Pile of Records into a Retrospective You Can Present
Three stages in one assignment — reproduce a full GRAI run on the course materials, run the whole thing on a real project of your own with a time comparison, then fix it into your team's retrospective SOP. All three, or it does not count.
Assignment goal
Connect the three things from the lesson and hand in something you can take straight into your next retrospective:
- Reproduce: run all three prompts on the course materials ("AI Open Class · Back-to-School Special") in your own AI tool;
- Apply: pick a real project of your own (even a single event or one month of work), assemble the five material types using the material-check template, and run the full process;
- Advance: turn the process into your own team's retrospective SOP — fixed material checklist, fixed prompts, fixed output format.
The score is not about how polished the report looks. It is about every root cause you write being traceable to a specific source.
Requirements
Hand in exactly three things: a full GRAI output on the course materials (comparison table + three-layer root causes + action plan), a retrospective record for a real project of yours, and your own retrospective SOP. All three, or it does not count.
- Use a real project: part one uses the course materials, part two uses a project you actually ran — the case from the lesson does not count.
- Mark information gaps: keep the "information gap" label for data that is not in the materials; do not fill in a number yourself.
- Root causes may not be people: this is the hard rule of this course — anything you could write as "someone was not careful enough" gets rewritten.
- Format: one document, or paste directly into the submission area; within three screens, no long screenshots.
- Deadline: parts one and two within 48 hours of the lesson; part three (the SOP) can follow within 7 days, since it needs a real run.
Review the four red lines first: every figure must trace back to the source data sheet; anonymise before feeding a real project (client names, internal pricing, people's names become placeholders); AI output is a first draft — review it yourself before it goes out; retrospectives target processes, not people — root causes point at processes and mechanisms, and AI output must never become a case file against a colleague.
Part 1: Reproduce (full GRAI on the course materials)
Run all three stages on the five materials (two sets of minutes, the chat excerpt, the task list, the activity data) and paste the output as it came.
| Item | My answer |
|---|---|
| AI tool I used | |
| Run A: how many events did the timeline restore | |
| Run A: how many contradictions did AI flag, and which was the most important | |
| Run A: how many information gaps did AI list | |
| Run B: the goal chain as written out (copy it, in the form A×B×C＝D) | |
| Run B: which funnel stage had the lowest achievement rate, and what deviation clue was given | |
| Run B: did AI flag the 28 conversions recovered by hand between 9/23 and 9/27 separately | |
| Run C: how many root causes did the analysis finally converge on (not four parallel causes) | |
| Run C: were the four delayed items identified as one chain | |
| Run C: how many actions, and how many lack acceptable criteria | |
| The question I followed up with + AI's answer |
Rows 3, 8 and 9 are the key observation points for this exercise.
- Row 3: the most hidden item in the materials is this — the task list shows "design requirement document" as complete, noted "on time", while the group chat shows that email contained only the headline, with no size list and no sub-copy. Did AI catch this kind of "on-time status, substandard content" implicit contradiction? If it did, your "flag contradictions" instruction worked. If not, feed the chat and the task list together again and require it to "cross-check how the same event is described in different materials along the timeline".
- Rows 8 and 9: if AI gives you four parallel delay causes (poster rework, article shifted, promotion shifted, referral compressed), the instruction "if several problems converge on the same root cause, say so explicitly" was ignored — add template 4's last line and run it again, then paste the second result too.
Self-check
- Every timeline row carries a source file, and three random rows can be found verbatim in the materials?
- Are the information gaps specific items, not one vague "materials incomplete"?
- Do the achievement rates match the activity data table?
- Does any root-cause cell contain a person's name? If so, rewrite it.
- Ask it: "Are the four delayed items one chain?" — see whether it can stitch the chain together itself.
Part 2: Apply (run it on a real project of yours)
Pick a project you actually ran. It does not matter if the materials are incomplete — reconstruct a lightweight version of whatever is missing (no group chat? list the key milestones by date), but state clearly in the table what you reconstructed.
| Item | My answer |
|---|---|
| Project + one-sentence description | |
| Time span | |
| Material completeness (goals / minutes / chats / task list / data — which you have) | |
| How I reconstructed what was missing | |
| Which templates I used | |
| How long the whole process took | |
| How long I estimate rebuilding the timeline by hand would take | |
| The original goals and measurement basis (quote the source) | |
| Actual results and achievement rates | |
| The root cause I dug out (write it out; it should point at a process or a rule) | |
| What I changed in AI's report (be specific) | |
| When the first action item will start |
Do not write "polished the wording" for the changes row. Be specific: "changed the root cause from 'cross-team coordination was poor' to 'deliverables had no acceptance checklist'", or "restored the achievement rate to the formula in the kickoff minutes".
Those changes are exactly what you should add to your prompt's context and constraints next time.
Part 3: Advance (fix the process into an SOP)
Making this faster every time depends not on memory but on four fixed things.
| Item | My answer |
|---|---|
| The scenario I will fix (which kind of project / how often) | |
| My fixed material checklist (where each of the five types lives) | |
| The fixed prompts I will use (which of the seven, adapted as needed) | |
| My fixed output format (how many report sections, slide pages, canvas or not) | |
| My fixed cadence (retrospective within how many days of the project ending, who compiles the data) | |
| Where I stored this SOP | |
| How I will make sure it actually gets used (who reminds, which step it follows) |
You do not need a knowledge base to start. Two things are enough: save the chosen prompts in your usual AI tool (in custom instructions or an agent if available), and file each round's materials into one folder named "project + date". Using it three times beats building something elaborate once and forgetting it.
Grading
| Level | Standard |
|---|---|
| ⭐⭐⭐ Excellent | All three parts; correctly caught the implicit contradiction (on-time status, substandard content); root causes converged at process level and trace back to sources; specific changes on the real project and which prompt line they were added to; the SOP states where it lives and how it gets used |
| ⭐⭐ Good | All three parts; GRAI run end to end with a complete comparison table and root-cause analysis; a quantifiable time comparison on the real project; the SOP broadly takes shape |
| ⭐ Entry | All three parts; the three stages run on the course materials, producing a readable retrospective report |
The two usual reasons for missing ⭐⭐⭐: root causes stopped at "execution was not solid enough / communication was not timely" (all true, and next round hits the same pit), and action items without acceptance criteria — the items are complete but no column answers "how do we know it is done".
Submission
- Upload a screenshot of each finished deliverable to the website to submit.
- Pick the one you are happiest with: post it in the course group so everyone can see your workflow.
One last reminder: before the report goes out, check every figure and every root cause once more. Figures must trace back to the source data sheet, root causes must trace back to specific material — these are the two places AI is most likely to err, and least allowed to.