Position in Module#
- Unit: 2
- Topic: From Structuring to Alignment: Clarifying Before Generating
- Literacy perspective: Interaction, secondarily Human-Context
- Collaboration level: Alignment
- Fundamentals focus: Generative AI Fundamentals
Literacy Perspective#
- Primary literacy: Interaction
- Secondary literacy, if any: Human-Context
- What this literacy explains: Clarification changes the input before generation. The human resolves consequential ambiguity and decides whether the proposed task interpretation is acceptable.
- What participants should manipulate, observe, and interpret: Compare direct LLM results with local Python results for the same analysis. Examine assumptions, accuracy, traceability, reproducibility and whether saved decisions govern later writing.
Collaboration Level#
- Level relation: Alignment
- Collaboration move being learned: Ask what is missing or undecided, resolve it, and approve a written basis for execution.
- What participants should be able to do differently after this Friday: Continue work from checked Markdown files with less dependence on conversation history or a particular interface.
Keep Unit 1's five phases: context, clarification, approach, execution and review. Deepen clarification and persistent working state. Distinguish two environments:
| Environment | What it provides | What to demonstrate |
|---|---|---|
| HAIC Chat | A virtual Markdown workspace shared across conversations | Save an output; create and edit a file; insert it in a later prompt. It does not expose a host file system or execute code. |
| VS Code with an AI coding agent | File access and execution in a bounded local folder | Specify, generate, inspect and run Python analysis. Rerun the retained code without another model call. |
This previews System Literacy. It does not require custom agent architectures or automation. Humans approve the moves from clarification to planning and execution.
Generative AI Fundamentals#
Explain why alignment changes generation: it changes the context from which the model continues. That context may contain dialogue, examples, constraints, decisions and files actually read by the system.
Use an ambiguous word such as Bank: a park context and a finance context invite different interpretations. Transfer this to a task whose audience or success criteria are missing. The language-play example can support this discussion; detailed tokenisation belongs to Unit 3.
A saved file is persistent state, but it only influences a model when supplied or read. Demonstrate a fresh session recovering the task from files. Also show a missed or weakly followed constraint: long context does not guarantee that every earlier instruction governs the result. Alignment improves the basis for generation; it does not verify the output.
Recurring Agentic Coding Demonstration#
Use one small public tabular dataset and one approved analysis specification. First obtain a CSV and figure directly from an LLM. In a fresh context, have the agent create Python that performs the same analysis. Run and rerun it locally; keep code, inputs, tests and outputs inspectable.
Compare both result sets against the specification. Correct the Python implementation where needed without weakening the criteria. Show that reproducibility and correctness require separate checks: rerunning the same code can reproduce the same mistake.
Then use only the verified Python results for report planning and writing. The direct LLM results and comparison remain in the evidence dossier. Use the challenge's file names so the demonstration transfers directly.
Friday Activity Table#
| Title | Content and activity | Observable output | Tools |
|---|---|---|---|
| Reusable Markdown state | Save and edit files in HAIC Chat; resume in another conversation. Distinguish virtual storage from local execution. | A task recovered from a file | HAIC Chat |
| Clarify before generating | Use the guest material to uncover assumptions, define success, ask questions and set scope. Distinguish task clarification from testing a product with users. | Human decisions and accepted task basis | Guest slides, alignment checklist |
| Set up local work | Install and verify VS Code, the coding agent and Python; explain the bounded folder and permissions. | A working local test program | Setup guide: Markdown, PDF; peer support |
| Build analysis that runs without AI | Specify a small analysis, generate Python, inspect outputs and rerun it. Compare with direct LLM results. | CSV, figure, code and comparison evidence | Dataset; WAILAI guide: Markdown, PDF |
| Challenge start | Explain the simulated report, nested workflows and A–E file handoffs. Start the question, data and common analysis specification. | Draft project and analysis files | Challenge 2 |
Experiment Focus#
- Experiment focus: direct statistical results from an LLM versus results from AI-authored Python executed locally.
- What changes: the method of producing the results and the inspectable code retained in B.
- What remains constant: approved
ANALYSIS.md, dataset and required result types. Both conditions produce at least a CSV table and PNG figure.
Keep the direct results once both files exist. Python may be corrected against the original criteria. Do not make code elegance an assessment target.
Observation Focus#
- Observation focus: numerical agreement, included records, missing values, assumptions, calculations, figure quality and reproducibility; then report fidelity to approved decisions.
- Evidence participants should capture: both result sets; Python specifications, code and tests;
evidence/ANALYSIS_COMPARISON.md; human decisions; report brief and plan; final report and dossier. - Interpretation participants should be prepared to make: Explain differences or justified agreements, what could be reproduced, which decisions needed human judgement, what a new context recovered and what remains uncertain.
Challenge Kickoff Deliverables#
Participants leave with a manageable documented dataset, a working local setup or agreed fallback, and a bounded question requiring at least one analysis, a CSV and a figure. Start .agents/PROJECT.md and .agents/ANALYSIS.md, note data limitations, and predict likely differences between the two methods.
Explain these handoffs before participants leave:
| Context | Inputs and purpose |
|---|---|
| A | ANALYSIS.md and data → direct results in results/llm/ |
| B | Same analysis and data, plus PYTHON.md and CODE_PLAN.md → code and results/python/ |
| C | Analysis and both result sets → evidence/ANALYSIS_COMPARISON.md |
| D | PROJECT.md, ANALYSIS.md, PYTHON_INTERPRETATION.md, results/python/ → approved PAPER_BRIEF.md and PAPER_PLAN.md |
| E | D's report basis plus brief and plan, without earlier dialogue → paper.md and Übungsbericht.pdf |
The code plan and report plan have different roles. The comparison is excluded from D and E. The required submission remains the report, dossier and complete work folder in one ZIP; the challenge defines the details and permitted professional-document alternative.
Teacher Setup#
Prepare working HAIC Chat access, the guest slides and checklist, the local setup guide, a public dataset and a tested Python example. Have prepared outputs and installation support available. Keep the project folder small and explain what the agent may read, change and execute.
Use the challenge folder structure and file names. Prepare an example of a fresh context recovering the task from the approved files. Choose the coding agent, script or notebook, work mode, support depth and additional handoff practice to suit the group. Keep the course path in VS Code and local Python.