systems-ai-workshop

Build a single self-contained HTML mini-tutorial that teaches ONE concept through a GENUINELY interactive centerpiece — something the reader manipulates that responds live — then verify it actually works before handing it back. The *form* of the interactive is chosen to fit the concept (a network process, a parameter model with sliders, an agent-based grid, a step-through algorithm, a data explorer, a what-if calculator, …) — contagion-on-a-network is just one example, not the default. Every build opens with a big-picture question, puts a hands-on interactive at the center with live-updating readouts, uses a fixed color scheme and uniform headings, and ends with tap-to-reveal learning checks (phone-friendly). Use when a faculty member wants an interactive explainer, a lesson dashboard, a concept simulator, or a "mini tutorial" they can open in a browser or share on a phone.

A monthly routine for scholars: search the web for recent mentions, press, talks, and citations of a named researcher and their work, then append the genuinely-new ones to a CSV (date logged, date published, title, source, URL, one-line description) — deduped against what's already there. Use when a researcher wants to track who is talking about them and their work over time, with links they can skim later.

Have the agent propose a concrete, numbered plan and get approval BEFORE executing — then work the plan step by step, re-planning out loud when reality diverges. Part of the "keep the agent on track" suite. Use for any task big enough that a wrong first move is costly, or when an agent tends to charge ahead.

Maintain a persistent house-rules file (CLAUDE.md or AGENTS.md) the agent reads every session — what the project is, how to run/test it, conventions, architecture pointers, and gotchas/"don'ts" — so it stops relearning the project and repeating mistakes. Part of the "keep the agent on track" suite. Use for any repo an agent will touch more than once.

Break a complex build into thin, end-to-end slices — each independently runnable and verifiable — instead of horizontal layers, so you always have something working and learn early. Part of the "keep the agent on track" suite. Use when a task is big enough that "build it all, then test" risks weeks of invisible drift.

Keep an append-only log of decisions (Architecture Decision Records) — each with context, the decision, and its rationale and consequences — so the agent doesn't re-litigate or silently contradict past choices. Part of the "keep the agent on track" suite. Use when a project accumulates choices you'll need to remember the "why" for.

Pin down checkable success criteria (and ideally tests) BEFORE building, so "done" is objective, not vibes — acceptance criteria, test cases, edge cases, and a done checklist the agent must satisfy and report against. Part of the "keep the agent on track" suite. Use to keep an agent honest about whether a task is actually complete.

Write a short design doc (RFC) before building — problem, goals & non-goals, proposed approach, alternatives considered, risks, and a validation plan — and get sign-off before the agent writes code. Part of the "keep the agent on track" suite. Use to stop an agent from charging ahead on the wrong solution.

Maintain a running handoff document with a checkbox task list the agent updates every session — current state, what's done, what's next, open questions, and a dated log — so work survives context resets and multiple sessions. Part of the "keep the agent on track" suite. Use for long, multi-session tasks or when handing work between people or agents.

Have the agent encode project structure as Mermaid diagrams — architecture, data flow, module dependencies, entity relationships, and state — so agent and humans share an accurate mental model of how things relate. Part of the "keep the agent on track" suite. Use when a system has enough moving parts that prose alone loses the relationships.

Maintain a living requirements doc that anchors an agent to system & subsystem goals, user needs, functions, constraints, and acceptance criteria — with traceability from need → function → test. Part of the "keep the agent on track" suite. Use at the start of a complex build, or when the agent keeps drifting from what the system is actually for.

Guidance for coding effectively with a LOCAL model (Ollama via the Cline extension) instead of a frontier cloud model — scope tasks small, lean on git for safety, choose the right local model for the hardware, raise the context window for agents, and verify output under zero-trust. Use when the user is working offline/privately with Ollama, asks why a local model is weaker, or wants the best workflow for local AI coding on a laptop.

Run an adversarial peer review of the user's own paper, grant, or research argument — surfacing the strongest objection, auditing claims against evidence, scrutinizing methods, flagging logical gaps, and returning a prioritized, actionable revision list. Use when the user wants to pressure-test a draft, simulate a hostile reviewer, respond to reviewer comments, or strengthen a research claim before submission.

Help the user turn a recurring academic chore into a reusable, reliable routine (a saved Skill): clarify the task and what good output looks like, draft a SKILL.md with clear trigger conditions and step-by-step instructions, then tighten it against a real example. Use when the user wants to automate or standardize a repetitive task — emails, summaries, letter outlines, meeting-notes-to-action-items, feedback synthesis — instead of re-prompting.

Generate clean, self-contained, print-ready HTML documents — course flyers, research posters, departmental memos, event one-pagers — with inline CSS sized for US Letter or A4. Use when the user wants a polished printable handout, flyer, poster, or memo they can open in a browser and save as PDF, rather than fighting with Word.