Agents & MCP

1. 1Block 1 hour on your team calendar (everyone must attend)
2. 2Each person opens their favorite AI tool (Claude, ChatGPT, Copilot, Gemini, etc.)
3. 3Identify a task where AI is underperforming or you want deeper understanding
4. 4Use this prompt: "I am having problems with [X]. I really want to [Y]. Interview me and ask a varied set of questions until you have 95% understanding of my problem, goal, and nuance — then work to solve it, or save it as a context doc I can refer to later"
5. 5Dictate back-and-forth with AI (optionally use Wispr Flow or similar for voice dictation)
6. 6Build context docs for: personal goals, 2026 team goals, family, friends, team members, vendors
7. 7Reuse these context docs in future AI sessions for better, more personalized outputs

Use Vercel Agent Browsers instead of screenshot-based tools like Claude for Chrome or Playwright MCP when building browser automation for AI agents. It provides Claude with a clean accessibility file and smart references, cutting token costs by up to 15x per screenshot.

1. 1Define metrics for what 'good' output looks like for your specific feature.
2. 2Use an evaluation framework to automatically grade model outputs (can use another LLM as a judge).
3. 3Analyze statistical trends in evaluations over time and across many samples.
4. 4Use the framework's sample synthesis feature to generate additional test cases using language models.
5. 5Integrate evaluations directly into your test suite.

Find the Kanban board project by @om_patel5 on GitHub or their profile. Set up the board, create task cards for your coding tasks, and drag them into 'In Progress' to trigger Claude Code automatically. Monitor multiple parallel agent sessions from the board.

Watch the full 24-minute Boris Cherny talk for the 40 prompting techniques. This 2-min clip is just the intro. Find the full video via @aiforgez's profile or caption link.

1. 1Before adopting any new AI model in enterprise: run a full evaluation for compliance, cost, and security standards — never just swap in the latest release.
2. 2Give developers control levers — custom modes, custom RAG systems, configurable tool-calling — so they decide how much agency the AI has.
3. 3Run vulnerability scanners during code generation, but avoid alert fatigue: popping up every 2 seconds causes developers to turn it off entirely.
4. 4Prioritize UX trust: if the AI keeps interrupting or overclaims credit, developers stop trusting it. Design so engineers feel they did the work alongside the AI.
5. 5Remember only ~1/3 of productivity gains come from the model — good architecture and engineering matter just as much.
6. 6AI is best framed as a boost (like better equipment for athletes), not a replacement: 60% of software workload is modernization/migration, and that work still requires skilled engineers.

1. 1Set up a Builder + Checker agent loop in Claude Code:
2. 2Create a Builder agent: its only job is to write and fix code.
3. 3Create a Checker agent: its only job is to run tests, type checks, and linting, then report exactly what failed.
4. 4Keep them separate — the same agent that writes code cannot reliably check its own work.
5. 5Create an orchestrator command that chains them: Builder implements → Checker runs → if all green, stop; if failures, send back to Builder → repeat.
6. 6Add stop rules: (a) max 5 cycles — if not fixed after 5, a human must intervene; (b) stop immediately if a fix breaks something previously passing (regression detected).
7. 7Without stop rules, agents will start weakening tests to fake a pass.

1. 1When setting up an AI loop:
2. 2First ask: can the goal be measured with a concrete value? (load time, error count, coverage %) → Use a deterministic loop with a numeric target Claude checks itself.
3. 3If no clean metric exists (code quality, doc accuracy, prose clarity) → Use an LLM-as-judge loop, but write a detailed rubric for Claude to test its work against.
4. 4Default rule: try to reduce the goal to something deterministic first. Only hand judgment to the AI when it genuinely can't be measured.

1. 1Design AI agent systems with end-to-end accountability:
2. 2Give the agent an independent identity, not just a user identity.
3. 3Define explicit permission boundaries: what the agent can and cannot do.
4. 4Use dynamic scoped policies — grant only the access needed for each specific task.
5. 5Log every tool call, script run, database change, and failure back to the specific agent and human owner.
6. 6Express constraints as intent-based rules (e.g. 'do not cause reduction impact') not just action blocks — agents will find alternate paths around simple 'no' rules.
7. 7Have humans design the goal, permission boundary, escalation point, audit trail, and accountability chain — not approve every individual action.

1. 1Identify exactly what would be installed: the free, open-source Clips Chrome extension by steve8708 (records screen + audio + browser logs/network into one shareable link). Do NOT install it yet.
2. 2Research it first — what it does, that it's open source (the repo can be reviewed), who publishes/maintains it, and what browser permissions it requests (it captures network requests and client errors, and redacts sensitive info). See Find the Resource for the repo.
3. 3Present the findings for review, and install ONLY after explicit approval.
4. 4Once approved and installed: from any tab, hit record, talk through/demonstrate the bug, then copy the generated link and paste it to the agent — the agent can read the transcript, view video snapshots, and inspect the logs from the URL, with no extra plugin needed.

review visuals — audio transcript only captures the speaker introduction. The 5 key principles from the caption (Interactive Extraction, token efficiency, etc.) are likely displayed as text overlays on screen. Watch the video to read and capture these steps.