Summary#
Module 1 develops adaptive mastery in human-AI collaboration: predicting behaviour, designing interventions, checking results and adapting the approach as tools and conditions change. Nine Friday/challenge cycles build evidence for reliable, responsible work.
The guiding question is:
How do we design human-AI collaboration so that reliable, responsible outcomes emerge?
Overview and Structure#
Each Friday combines conceptual input, guided experiments and challenge preparation. A recurring coding example supports the teaching where useful. Saturday provides coaching and feedback on the challenge.
Three components guide the learning:
- Literacies identify dimensions of AI behaviour and control.
- Collaboration levels describe ways of working with AI.
- Generative and agentic AI fundamentals help explain observed behaviour and limits.
Units 1 and 2 introduce five supplied workflow phases: context, clarification, approach, execution and review. Unit 1 uses browser conversations without file access. Unit 2 carries decisions between sessions in Markdown files, then uses VS Code and local Python for reproducible analysis. These are first encounters with the whole process. Unit 2 deepens clarification, Unit 5 planning and execution through co-construction, and Unit 6 review and improvement. The remaining units develop mechanisms, grounding, human responsibility, and orchestration.
Each challenge contributes predictions, outputs, observations, interpretations and practical rules to an evidence portfolio. Participants draw on this portfolio in Module 2.
How Levels, Literacies, and Fundamentals Work Together#
The five literacies answer: what needs to be observed or controlled?
- Interaction: requests, exchanges, examples, clarification and feedback.
- Mechanism: tokens, attention, parameters, sampling and variability.
- Data: sources, grounding, embeddings, retrieval, factuality and traceability.
- System: workflows, handoffs, checkpoints, tools, memory and orchestration.
- Human-Context: trust, responsibility, legitimacy, ethics, leadership and social fit.
The collaboration levels answer: how are human and AI working together?
- Level 0: asking for answers.
- Level 1: instruction and structured requests.
- Level 2: alignment and clarification before generation.
- Level 3: co-construction and process design.
- Level 4: reflection, critique, revision and responsible judgement.
- Level 5: orchestration across steps, tools, roles and systems.
The fundamentals explain how training, token sequences, context, sampling, grounding and tools can affect behaviour. Practical demonstrations make limits, bias, instability and risk observable. Safeguards require controls over data, permissions, validation, actions and human decisions.
Keep these roles distinct. A level describes the collaboration mode; a literacy supplies the lens; a fundamental supports interpretation. Observed outputs alone do not reveal a model's internal process.
Recurring Reference Application#
AI-supported coding provides an instructor demonstration where files, tests and changes make decisions visible. It supports transfer to participants' own domains; programming is not the course's main subject.
Unit 1 stays in a browser without file access. Its main classroom example is a five-minute presentation; any coding illustration remains conversational. Unit 2 first introduces a virtual Markdown workspace in HAIC Chat, then a bounded local folder in VS Code. Specifications, plans, code, results and review notes preserve state across sessions. Local Python can rerun the analysis without another model call; reproducibility alone does not establish correctness.
The reference case can show:
- persistent context and an agreed specification before implementation;
- planning, execution, tests, review and approval as separate stages;
- evidence from outputs, code changes, tests and review findings;
- delegation through roles, subagents or separate workspaces;
- grounding in files, documentation and other checked sources;
- safety and security boundaries: source separation, limited data and permissions, validation, approval of consequential actions, audit records and accountable human oversight.
Use the case when it makes the learning move easier to inspect. Choose tools and infrastructure to serve that purpose.
Course Structure (Fridays and Challenges)#
| # | Friday Topic | Literacy | Level | Fundamental | Friday Demonstration or Reference Case | Challenge |
|---|---|---|---|---|---|---|
| 1 | From Asking to Structuring: Designing AI Interaction | Interaction | Instruction | Next-token prediction and variable output | Direct request versus the supplied five phases; five-minute presentation in a browser without file access | Five-Phase Collaboration Experiment |
| 2 | From Structuring to Alignment: Clarifying Before Generating | Interaction; Human-Context and early System perspective | Alignment | Context, persistent state and reproducible local execution | Reusable Markdown in HAIC Chat; direct LLM analysis versus local AI-authored Python in VS Code | Reproducible Alignment Experiment |
| 3 | Why Models Behave Differently: Parameters, Variability, and Control | Mechanism | Instruction with mechanistic control | Tokens, attention, sampling and variability | Repeated trials of one task under two controlled conditions; optional coding comparison | Parameter Control Experiment |
| 4 | From Opinion to Evidence: Grounding AI in Data and Sources | Data | Instruction with grounded evidence control | Embeddings, vectorisation, retrieval, grounding, and bias | Coding with minimal context versus grounded repo/API/test context | Grounded Generation Experiment |
| 5 | From Answers to Processes: Designing Human-AI Co-Construction | System, with Interaction | Co-construction | Decomposition, project context, persistent state and reusable skills | Equipment loans: three cuts, artefacts, checkpoints, perspectives and roles; evaluate quality, control and effort; PDCA | Co-Construction Workflow |
| 6 | Reflection Loops: Critique, Revision, and Better Reasoning | Interaction; Mechanism | Reflection | Reflection workflows, evidence for human judgement and LLM-as-judge | Three lessons on nested reflection workflows; six Arena lessons: inspect data, improve the judge, then improve the generator | Reflective Improvement Workflow |
| 7 | Leadership, Responsibility, and Legitimation in Human-AI Systems | Human-Context | Reflection with responsible co-construction | Fluent generation versus legitimate decision support; substantive oversight, contestability, and calibrated reliance | Consequential coding proposal with varied authority, evidence, approval, and escalation logic | Responsible Decision Logic Experiment |
| 8 | Designing Workflows: From Single Chat to Reliable AI Systems | System | Orchestration | Agentic AI: perception, reasoning, planning, tools, reusable capabilities, workflow state, and control points | Coding agent using persistent context, tools, resources, permissions, validation, and approvals | Workflow Control |
| 9 | Security and Trust Boundaries in Human-AI Collaboration | System, Data, and Human-Context, with Interaction as the attack surface | Orchestration with security-aware control | Everyday generative and agentic AI security: OWASP-informed risk mapping, injection and jailbreaking, source/RAG and memory poisoning, sensitive-information exposure, layered safeguards, least agency, and residual risk | Office or agentic coding workflow exposed to manipulated content; compare weak boundaries with scoped data and tools, validation, exact-action approval, and audit traces | Security Control Experiment |
Unit 9 Scope#
Unit 9 concludes Module 1 by integrating the preceding literacies and collaboration levels into security-aware orchestration. Current OWASP guidance for LLM and agentic applications is used as a reference map rather than as a taxonomy participants must memorize.
The experiential focus is everyday office use: indirect prompt injection through documents or messages, weak or poisoned retrieval sources, sensitive-information handling, layered safeguards, tool permissions, agent capability, human approval, and appropriate reliance. Training-data poisoning, model backdoors, model inversion, membership inference, and specialist embedding attacks remain awareness topics. Operational safeguard bypass, unauthorized testing, exploit development, and extraction of real protected information are outside the unit scope.
Relation to Module 2#
Module 1 concludes with security-aware orchestration. In Module 2, participants select relevant portfolio evidence to redesign a work practice, develop a Human-AI Collaboration Blueprint and explain their personal theory of control.
Challenge Design Logic#
Each challenge makes a control problem testable in the participant's context. Its opening should state what to produce, what to compare, what stays fixed and what to submit.
The detailed brief then defines:
- Task: what is not yet controlled.
- Approach: the condition or design choice to compare.
- Artefacts: the work that another person can inspect.
- Evidence: the observations, traces or measurements supporting a claim.
- Interpretation: what the result implies for control, transfer and limits.
The challenges provide the main evidence of learning. A plausible output or an unsupported reflection is not enough.
Unit Coordination#
See units.md for compact Friday/challenge summaries. Each numbered folder contains the detailed teacher-facing friday.md and the authoritative challenge.md.