Here it is!
AI is most useful when it reduces friction and helps you think, draft, search, summarize, analyze, or prototype faster. It becomes risky when it quietly replaces judgment.
A good rule of thumb:
What you delegate, you must validate.
1. Start with useful skepticism
Several of you brought up one of the most important AI habits: do not confuse confidence with correctness.
Large language models can give polished answers, cite sources that do not really support the claim, miss context, misunderstand industry jargon, or analyze only part of the data while sounding authoritative. That does not make them useless. It means you need a review process.
When using tools like ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, or NotebookLM, ask the model to show its assumptions, separate facts from inference, provide reproducible steps, and point back to the source material. ChatGPT’s data analysis tools can work with uploaded data to create tables and charts, and NotebookLM is especially useful when you want the model grounded in documents you provide.
A practical prompt to reuse:
Before answering, list what you know from the provided source, what you are inferring, what you are uncertain about, and what I should verify manually.
Another useful prompt:
Give me the answer, then give me the skeptical reviewer version. Where could this be wrong?
2. Ground the model in your actual context
A recurring theme in our discussions was that generic AI becomes much more useful when it has the right context. That context might be a customer account history, a policy document, a job description, a compliance requirement, a fleet operations process, a hospitality workflow, a research paper, or a set of internal acronyms.
That is the idea behind RAG, or retrieval-augmented generation: instead of relying only on what the model learned during training, you give it access to relevant documents or data at the time of the task. LlamaIndex describes RAG as a way to add your data to what the model can use, while LangGraph focuses on building stateful agent workflows with concepts like memory and human-in-the-loop review.
Useful tools to explore:
NotebookLM — useful for working with a set of provided sources.
LlamaIndex — useful for building RAG systems and connecting LLMs to data.
LangChain and LangGraph — useful for building more structured workflows and agents.
Microsoft Copilot Studio — useful for organizations already working in the Microsoft ecosystem.
ChatGPT Apps and Connectors — useful for bringing tools and data into ChatGPT workflows.
The practical takeaway: the better the context, the better the output. But you still need to verify the output.
3. Use the “reverse prompt” technique
One of the most useful ideas discussed in class was asking the model to help you write a better prompt before asking it to solve the task.
Try this:
I need to extract useful information from this document for [purpose]. Before doing the task, write the best prompt I should use. The prompt should include: the role the model should play, the fields to extract, the format of the output, how to cite sources, how to label uncertainty, and what not to infer.
Then review the generated prompt, improve it, and use that version.
This works especially well for:
customer account summaries,
policy reviews,
research papers,
resume or candidate comparisons,
compliance documents,
project briefs,
meeting transcripts,
technical documentation,
sales or customer success preparation.
The goal is not to become a “prompt engineer” in a narrow sense. The goal is to become better at structuring work for AI systems.
4. Treat agents as junior teammates, not autonomous authorities
In the second live session, we spent a lot of time on agents. The useful way to think about an agent is:
model + objective + tools + memory/context + feedback loop.
An agent becomes more powerful when it can search, use tools, write code, open files, send messages, manage a calendar, update a database, or pass work to another agent. The OpenAI Agents SDK describes agents as applications that can plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work. Claude also supports tool use, and Claude Code is an example of an agentic coding tool that can read code, edit files, and run commands.
CrewAI focuses on multi-agent systems with guardrails, memory, knowledge, and observability; n8n’s AI Agent node connects models to external tools and APIs; Zapier Agents emphasizes automation across thousands of apps; and Make AI Agents emphasizes visibility into decisions, connections, and costs.
The safe starting point is not “let an agent run my life.” The safe starting point is:
draft but do not send,
summarize but do not decide,
recommend but do not execute,
prepare but do not approve,
log every action,
require human review at key steps.
5. Match the risk level to the use case
Some of the best use cases we discussed were practical and relatively low-risk:
meal planning and grocery lists,
travel planning,
customer meeting preparation,
research summaries,
first drafts of emails,
coding assistance,
training plans,
interview practice,
role-play with a customer or stakeholder persona,
summarizing policies or documents,
turning messy notes into a structured brief.
A meal-planning agent that suggests dinners, checks inventory, compares grocery stores, and creates a pickup order is a great learning project because the downside is manageable. A model making decisions about hiring, lending, public safety, military targeting, medical care, insurance adjudication, or regulated compliance is a very different risk category.
The more consequential the decision, the more you need:
clear permissions,
audit logs,
human review,
source citations,
test cases,
security review,
approved enterprise tools,
and a defined escalation path.
6. Respect privacy and workplace boundaries
A major theme in both sessions was tool access at work. Some of you have access only to approved enterprise systems. Some work with sensitive data. Some are in regulated or public-sector environments. That matters.
Do not upload customer data, employee data, regulated data, confidential business information, legal material, source code, or operationally sensitive information into personal AI accounts unless your organization explicitly allows it.
For workplace use, start with approved systems such as Microsoft 365 Copilot, ChatGPT Enterprise or Business, Claude for Work, or whatever your organization has reviewed and approved. Microsoft’s Copilot documentation emphasizes that organizational data access depends on existing permissions and enterprise controls, which is exactly why governance and access settings matter.
For personal experimentation, local tools can be useful:
Ollama is available on macOS, Windows, and Linux; LM Studio is positioned around running AI models locally and privately; and Open WebUI supports Ollama and OpenAI-compatible APIs for self-hosted workflows.
But “local” does not automatically mean “safe for work.” Local tools still need the right device, data policy, model policy, logging, access controls, and security review.
7. Be especially careful with agents that can act
Agents become riskier when they can use email, calendars, files, browsers, APIs, credentials, payment systems, or third-party plugins.
This is why security resources matter. The OWASP Top 10 for LLM Applications includes risks like prompt injection, insecure output handling, sensitive information disclosure, supply-chain vulnerabilities, and excessive agency. NIST’s Generative AI Profile is a cross-sector resource for organizations thinking about trustworthiness and risk management in generative AI systems.
For OpenClaw specifically, the official site describes it as a personal AI assistant that can clear inboxes, send emails, manage calendars, and check users in for flights; that kind of capability is powerful, but it also raises security concerns when connected to real accounts and third-party skills. OpenClaw announced VirusTotal scanning for its skill marketplace, and Cisco has warned that agents with system access can become covert data-leak channels without appropriate controls.
A practical safety checklist for agents:
Start with read-only access.
Use draft-only mode before send mode.
Do not give the agent primary credentials early.
Avoid financial, legal, medical, HR, or safety-critical actions at first.
Log what the agent saw, decided, and did.
Review third-party tools and plugins as if they were software installed on your machine.
Create a kill switch.
Never assume “the agent probably knows what it is doing.”
8. Keep logs and evaluate what works
One of the strongest ideas from the discussion was the importance of memory, logs, approvals, and feedback. If an agent or AI workflow produces ten outputs, you need a way to know which were useful, which were wrong, and why.
For personal use, that might be a simple spreadsheet:
task,
prompt,
tool used,
source material,
output,
what I accepted,
what I changed,
what was wrong,
what I would ask differently next time.
For teams building AI systems, look at tools like:
LangSmith provides workflows for defining datasets, evaluators, and experiments; Phoenix traces model calls, retrieval, tool use, and custom logic; Braintrust supports systematic AI evaluation and production monitoring; and OpenAI Evals supports programmatic evaluation of LLM applications.
The habit matters more than the tool: keep track of what you tried and whether it actually worked.
9. Use AI for role-play, but understand the limits
Several of you brought up using AI as a coach, mock interviewer, customer persona, stakeholder, or even a challenging debate partner. This can be extremely useful.
Try prompts like:
Act as a skeptical customer in [industry]. Challenge my recommendation and ask the hardest questions you would ask before approving this purchase.
Or:
Act as a hiring manager for [role]. Interview me for 15 minutes. After each answer, give me direct feedback and one stronger version of my response.
Or:
Act as a senior executive who is short on time. Review this recommendation and tell me what is unclear, unsupported, or too technical.
But remember: an AI persona is not a real customer, real regulator, real patient, real employee, or real citizen. It is a simulation. Use it to practice, not to replace actual discovery, research, or stakeholder feedback.
10. Explore multimodal and creative tools
We also discussed that AI is no longer only text. Voice, image, audio, video, and music are becoming part of the practical toolkit.
For music exploration, especially if you are trying to avoid lyrics or create instrumental ideas, try:
Suno Instrumental Generator
Udio
AIVA
Suno offers an instrumental generator for backing tracks, Udio is an AI music generation platform, AIVA supports downloads such as MIDI, WAV, stems, and MP3, and MuseScore can import MIDI files and convert them into music notation.
For sheet-music-oriented work, a practical workflow is:
generate or sketch the musical idea,
export MIDI if the tool supports it,
import the MIDI into MuseScore or another notation tool,
clean up the notation manually.
With AI music tools, also be mindful of copyright, licensing, and your intended use.
A simple exercise after these sessions
Pick one real task from your work or life and write down:
What am I trying to accomplish?
What data or context would help?
What tool is approved or safe for this context?
What is the lowest-risk version of this task?
What should the AI draft, summarize, search, or organize?
What must a human still approve?
How will I verify the output?
What should be logged for future improvement?
Then test it on a small, low-risk version of the problem.
For example:
Instead of “AI should manage my inbox,” start with “AI should summarize five non-sensitive emails and draft suggested replies.”
Instead of “AI should analyze all customer data,” start with “AI should help me prepare questions for one customer meeting using approved notes.”
Instead of “AI should build an autonomous workflow,” start with “AI should produce a checklist, then I manually execute it.”
Instead of “AI should make a decision,” start with “AI should identify options, risks, and missing information.”
That is how you build real capability without skipping judgment.
Final thought
The people who benefit most from AI will not necessarily be the people who chase every new tool. They will be the people who learn how to:
ask better questions,
provide better context,
verify outputs,
protect sensitive data,
design safer workflows,
use automation where it fits,
and KEEP HUMAN JUDGMENT IN THE LOOP!