Bee transcript · May 29, 2026
Working with AI agents.
Two tabs: the source transcript from the AI Lab session, and a cleaned summary of the practical lessons on context, tools, skills, verification, and QA workflows.
Core idea
Agents do not need magic prompts. They need access, boundaries, tools, and verification.
Best first move
Ask the agent what it can see before asking it to debug, browse, file bugs, or change code.
Safety rule
Never paste passwords into shared prompts or reports. Give the agent safe local access instead.
Mental model
The stack: interface → model → tools → skills → verification.
Context before instructions
The agent can only work with what it can actually see: the right folder, repo, page, browser session, task description, and expected output.
The interface, model, and tools are separate
Cursor, Codex, Claude, GPT, Hermes, and browser automation are different layers. If one setup stalls, the fix may be the model, the interface, or tool access — not more prompting.
Authenticated pages need real access
Being logged in as a person does not automatically mean the agent can inspect the same page. Private app testing needs safe local browser/tool setup, not copied passwords.
Tools are capabilities; skills are repeatable methods
A browser, terminal, GitHub, search, and file access are tools. A QA audit, bug report, research pass, or verification checklist is a skill.
Trust exploration, verify conclusions
Agents can find useful bugs, but they can also confidently report things that are not true. Good workflow means checking reproduction steps before sharing findings.
Parallel agents come after one clear workflow
Multiple agents help when each has a bounded job: QA, UX review, technical debugging, research, or report writing. Without boundaries, they create more noise.
Practical QA workflow
How to use an agent to test a web app without filing nonsense.
- 1Open the correct app, folder, repo, or page before asking for help.
- 2Tell the agent its role: QA tester, debugger, product reviewer, developer, or researcher.
- 3Specify the output: bug report, audit, test plan, reproduction steps, or implementation plan.
- 4Ask it to inspect core flows: signup/login, forms, navigation, buttons, error states, empty states, and mobile behavior.
- 5Require every issue to include steps to reproduce, expected result, actual result, severity, evidence, and verification status.
- 6Manually verify the top findings before filing or sharing them.
Session phrases
Keep these.
Не думай, что агент видит то же самое, что видишь ты.
Do not assume the agent sees what you see.
Инструмент — это что агент может использовать. Скилл — это как и когда это использовать.
A tool is what the agent can use. A skill is knowing when and how to use it.
Отдельный навык — понимать, когда модель ошибается.
A separate skill is learning when the model is wrong.
Prompt starters
Copy these into your agent.
Act as a QA tester. Inspect this web app and create a bug report. For every issue include page, steps to reproduce, expected result, actual result, severity, evidence, and whether it was verified. Do not invent bugs; mark uncertain items clearly.
Before you start, tell me what context you can access: files, browser, repo, page, terminal, search, or none. If you cannot see the app, say exactly what access is missing.
Review this agent's report. Separate verified issues from guesses, missing evidence, and follow-up checks.
Source note
Transcript and summary are separated.
The Transcript tab is the Bee-recorded AI Lab conversation from the verified 2–4 PM source window. The Summary tab is the edited lesson version. The unrelated post-session account/photo-storage tail is not included.