I can't sleep gud anymore - A Practical Guide to Agentic Computering

Mario Zechner demonstrates a control-heavy Claude Code workflow built around explicit context, state stored on disk, and human review. The most transferable pattern is to treat a prompt as a program and small Markdown or JSON artifacts as inspectable state that the agent queries and updates.

Practical Workflow

• Use agents for disposable automation, not only production code: Zechner asks Claude Code to convert media and write small Bash or Node scripts, then keeps only tools that become reusable 09:00–13:45(opens in a new tab).
• Externalize project context, then review it as ground truth: He has Claude draft project and module documentation, corrects omissions manually, and proposes updating the relevant document from the final diff after each change. He repeatedly warns that generated documentation is a first draft and can be wrong 14:40–25:55(opens in a new tab).
• Recover before compaction, not after it: When a session nears its context limit, he writes a task-summary file, compares it with the conversation, iterates until important decisions are retained, then starts a fresh session with the project spec and task summary. His demo also shows the first summary losing proposal details, so the method still depends on manual checking 47:30–59:50(opens in a new tab).
• Make large-codebase context queryable: For a cross-language Spine runtime port, Zechner generates symbol indexes with lsp-cli, queries only needed records with jq, and keeps a port plan, target-language conventions, and accumulated porting notes on disk. This limits the context loaded for each unit of work while preserving progress between sessions 1:02:00–1:12:14(opens in a new tab).
• Put the human inside the workflow: The port loop opens the reference diff and target files, pauses before each type, compiles the change, shows the resulting diff, and waits for Zechner to review or intervene before marking the item complete. The demonstrated item is small; the broader 68-file, 5,000-line and one-day-versus-two-weeks figures are his report without task logs, quality measures, or an independent baseline 1:12:14–1:20:19(opens in a new tab).
• Treat generated tests as drafts: Zechner says Claude may over-mock behavior, weaken assertions, or remove them while resolving failures. His recommendation is to give the loop executable feedback but manually verify that the tests still assert the intended behavior; the video provides examples from his experience rather than a measured failure rate 1:23:11–1:24:16(opens in a new tab).

Permission prompts are a separate safety decision

In this July 2025 demonstration, Zechner bypasses Claude Code's permission prompts on a machine he says contains nothing irreplaceable. He reports no incident in two months and briefly suggests a container. That experience does not establish the safety of skipping permissions on machines or accounts with valuable data or credentials 02:38–06:52(opens in a new tab).

Watch the full video on YouTube(opens in a new tab).