Browser chats are useful for quick questions. For real project work, an agent needs context, a plan, the right tools, and a way to check the result.
These notes cover how an agent works, how to keep project context under control, 19 questions and clarifications from the preserved recording fragment, and four prompts you can use straight away.
An agent can move faster through the work, but the manager still owns the problem, priorities, access, and final check.
An agent becomes useful when it can work inside an organised project, use permitted tools, observe the result, and correct its work.
You move the context between steps yourself: upload a file, copy the answer, open another tool, then return with the result. Useful for a question, fragile as a project system.
You provide a goal and a permitted working environment. The agent can inspect project files, propose a plan, use tools, check the result, and make a controlled correction.
The model is only one part. Reliable project work comes from the whole system around it.
This pattern works in agentic desktop apps, terminals, and IDE extensions that can access a project and permitted tools. See the official descriptions for Codex workflows and Claude Code.
A large context window does not replace working discipline. When research, specification, implementation, and unrelated questions share one chat, old decisions start competing with the current task.
Choose one recurring task with a clear input and a result you can verify. Give it a project folder and a fresh chat. Ask for a plan first, approve one small step, compare the result with the primary source, and save the workflow as a skill only after it proves useful.
Every workflow tonight had the same four parts. Use them for any task you hand to AI.
The recording covers only part of the workshop, so this is not the total for the whole evening. The questions are grouped by workflow, team context, adoption, tools, and automation.
Both are possible. Before using a ready-made skill, review its instructions, scripts, requested permissions, and external actions. A skill can hide actions that expose, alter, or delete your data, or affect your device. Test it in a limited environment on your own inputs, rules, and expected output before giving it real access.
Yes. Check the goal, constraints, unnecessary scope, and how the result will be verified.
Yes. Ask for alternatives, change the constraints, or edit the proposed plan directly.
Use version control for text and code. Give every meaningful version a clear purpose and test result.
Agree on ownership boundaries, a shared source of facts, and project rules stored with the work.
An agent can prepare a digest from agreed sources. A human should review changes, risks, conflicts, and links to primary data.
The agent can make the conflict and trade-offs visible. A person with the right authority makes the decision.
That depends on the company process. A tracker can remain the control system while the agent also uses specifications, rules, and acceptance criteria.
First structure the data and check its quality. Then AI can help visualise relationships and surface possible blockers.
Define the source, permissions, refresh frequency, and summary format for each data type. A connection alone does not guarantee freshness.
Common causes are vague use cases, unclear access rules, too little practice, and weak baseline AI literacy.
Start with the vendor's official documentation. Use it as the primary source for features, limits, and changes.
Professional communities are useful for real cases. Check their advice against official documentation and your company's rules.
No. VS Code is a workspace for files and tools. MCP is an open standard for connecting AI applications to external data and actions.
Use a stronger model for unfamiliar, complex design work. A lighter model can handle repeatable steps with a clear instruction and check.
First prove one real workflow in a tool you already have. Expand the stack only for a confirmed working need.
Steps can be connected and triggered by an event or schedule. Risky actions still need boundaries, logs, error handling, and a human check.
We discussed regular digests and monitoring new data against defined criteria. Start with one stable step, not the whole chain.
No. Technical possibility is not the same as an allowed integration. Check the official API, platform terms, and blocking risk.
Each prompt includes the input, the expected result, and the checks that keep the final decision with you.
Input: a meeting transcript from Zoom, Teams, Meet or any recorder.
What you'll get: decisions, owners, follow-up tasks and a follow-up message, ready for your review.
Human check before you send
Input: daily or weekly updates from your team (chat messages, stand-up notes), plus an export from Jira, Azure DevOps, Linear, Trello or a spreadsheet.
What you'll get: a one-page status report for a senior stakeholder and a shorter version for the team.
Human check before you send
Input: a rough idea in a few lines. Include the numbers you have and any decisions already made.
What you'll get: a one-page spec that a team, a client or a builder tool can work from.
Human check before you build
Input: the spec from Prompt 3, after your check.
What you'll get: a clickable prototype you can open in a browser and show to your team or a client.
Where to run it: in an AI agent like Claude Code or Codex, or in a prototype builder like Lovable, Bolt or Replit. In Claude.ai or ChatGPT, ask for "a single HTML file" and open it in the preview.
Human check before you share
Fix things in plain words:
AI Sandbox is a hands-on program for non-coders: 5 sessions where you build your own AI workflows and tools, with a mentor, from idea to a working prototype.