2026-11-01
Capture what you learned before AI
The retrospective is your own account.
No AI drafting: it should reflect what you actually saw.
What did you plan in A1 and A2?
What did you actually build in A3?
Which design decisions made the code easier to write or test?
Name one, and point to the class or file.
Which decisions caused rework, awkward tests, or workarounds?
Name one, and say why.
Knowing what you know now, what would you change in A2?
Keep it to one concrete change.
Record the answers in your A3 retrospective:
docs/a3-retrospective.md
Five minutes each, then share one point with another team.
Weeks 1 to 8: manual design.
From now: AI-assisted revision.
Due Nov 16/17, midnight of your lab day.
Incorporate the change request:
| Section | Required content |
|---|---|
| AI critique setup | The prompt used to ask AI to critique A1 |
| Critique table | At least six suggestions, each accepted, rejected, or revised |
| Revised scope | Updated problem boundary, including any new out-of-scope item |
| Section | Required content |
|---|---|
| Revised requirements | Three changed or new user stories with acceptance criteria |
| Impact note | One paragraph on how the revision affects the A2 design and A3 code |
| Section | Required content |
|---|---|
| Design delta | What changes from A2 and what stays stable |
| AI alternatives | Two AI-suggested design alternatives, summarized fairly |
| Human decision | The chosen design, and why it fits this project |
| Updated diagram | One updated class, component, state, or sequence diagram |
| Section | Required content |
|---|---|
| Risk register | Three risks: design, testing, deployment |
| Claim verification | Check one consequential AI claim against authoritative documentation |
| A6 implementation plan | The smallest set of changes for the final prototype |
The quality of comparison and justification.
The point is to use your judgment of the AI’s response, not just parrot what it says.
docs/ai-log.md, from the first prompt on.
For each interaction record:
Write down, in the repo:
Give the agent your A1 and ask for:
Here is my requirements specification (A1). Critique it as a skeptical reviewer. List gaps, ambiguous wording, conflicting requirements, and untestable acceptance criteria. For each, quote the line and explain the problem. Do not rewrite it.
Adapt this prompt as needed to your project.
Every suggestion gets a verdict:
accepted, rejected, or revised
and a reason.
| AI suggestion | Verdict | Reason |
|---|---|---|
| “Add user login and roles” | Rejected | Out of scope for A1; no stakeholder asked for it |
| “‘Fast’ search is vague; give a limit” | Accepted | Replaced with: results appear within 2 seconds |
Agents state wrong things confidently.
Pick one consequential claim and check it against authoritative documentation.
Record the evidence and any correction.
docs/ai-log.mddocs/ai-log.md
Course home · Neil Ernst ©️