Podcast Wiki

Coding-Agent Workflow (practical setup)

5 min read

Summary: Day-to-day practices for working with coding agents (Claude Code, Codex, Cursor, Pi). Spec-driven development has become the consensus starting point (GitHub Spec Kit, OpenSpec, Superpowers), often with a “grill me” skill that interviews you before building and logs the assumptions you’re making. Keep the workflow harness-agnostic with a plugin like Superpowers (spec → plan → implement), so you can hop between tools as models and pricing change. Plan visually and at the right altitude, and give agents dedicated design guidance (e.g. Impeccable) since default AI-generated UI is weak. Tune harness settings, and build small personal tools and skills for the friction you hit: losing track across many terminals, vague agent questions, the laptop sleeping mid-run.

Key ideas

  • Harness hopping: one host moved to Codex after an enterprise pricing change and found GPT 5.5 stronger than Opus for long autonomous tasks, though slower (langtalks—70-our-claude-code-tips).
  • Superpowers’ spec/plan/implement flow with review subagents makes switching between Cursor, Claude Code, and Codex nearly free (langtalks—70-our-claude-code-tips).
  • Plan altitude: overly verbose technical plans blend the human’s requirements with the agent’s assumptions, so implementation treats them all as equal. Keep technical plans high-level and render them as HTML (langtalks—70-our-claude-code-tips).
  • Feedback tooling: a small server plus a Chrome extension let him highlight and comment on HTML plans, turning comments into live tasks. Feedback is often a principle (“never more than two buttons per page”), not a one-off edit (langtalks—70-our-claude-code-tips).
  • Claude Code settings he uses: no-flicker mode; background isolation off (no auto worktrees when supervising); his own git instructions instead of the default; skill-listing budget 0.02 → 0.03 (truncated descriptions hurt); cleanup period 9999 days (keep sessions); remote control on startup; knowledge-work plugins marketplace (langtalks—70-our-claude-code-tips).
  • Output style “context recap”: every reply ends with task (one line), what’s done, numbered next steps (reply “1” to continue), and open threads, so you never scroll back across parallel terminals (langtalks—70-our-claude-code-tips).
  • Skills: “ask good questions” (product-level questions, not obscure specifics); team “linear ops” (ticket ↔ session ↔ branch ↔ PR conventions) (langtalks—70-our-claude-code-tips).
  • Terminal: Warp with a tab per project and a pane per task, so context switches happen per project (langtalks—70-our-claude-code-tips).
  • Use idle subscription windows at night for routines: deep code and security reviews, Linear ticket dedupe (langtalks—70-our-claude-code-tips).
  • A keep-awake tool lets Claude Code keep running with the MacBook lid closed (handling heat and hotspot) (langtalks—70-our-claude-code-tips).
  • The scarce skill is managing your own attention across parallel agents, not the agent’s context: you must move up from bits and bytes to the bigger picture while agents work (langtalks—63-wake-up).
  • BMAD (a spec-driven framework) is overkill end to end, but its early analyst stage (Five Whys, Six Thinking Hats) surfaces product angles before code (langtalks—62-ai-rd-rollout).
  • Not every heavy user wants orchestration UIs: Idan runs 6–7 terminal tabs of Claude Code side by side and skips worktree-based tools (ignore-instructions—20-idan-benyon-build-ship-grow, idan-benyon).
  • The fun of Claude Code produces “project inflation”: many half-finished side projects, feeling spread too thin (ignore-instructions—20-idan-benyon-build-ship-grow).
  • Spec-driven development (SDD) is the hosts’ consensus, with GitHub Spec Kit, OpenSpec and Superpowers as the frameworks named (developers-outside-the-box—185-lets-talk-about-our-tools).
  • A “grill me” skill that fires many questions at you before building is “simply fantastic” for surfacing what you hadn’t thought through (developers-outside-the-box—185-lets-talk-about-our-tools).
  • For UI work: Claude Design, or a product-design skill used from another chat tool (developers-outside-the-box—185-lets-talk-about-our-tools).
  • Grill Me (Matt Pocock) interviews you in rounds and follows up on your answers. Use it when you’re most excited, which is when you skip awkward questions, and for small decisions too (annual plan? feature for one big client?). Add an instruction to log every “probably” (they probably have budget, probably gave access) as an assumption, then ask which to check first with a customer (startup-ideas—8-github-repos-viral-money, greg-isenberg).
  • Impeccable (Paul Bakaus) gives coding agents specific design guidance: set it up with who the product is for, then run commands like critique, polish, bolder and quieter. Start with critique on an existing page and fix only the parts you agree with (startup-ideas—8-github-repos-viral-money, greg-isenberg).

Disagreements & open questions

Takeaways

harness-engineering · skill-engineering · ai-verification

Graph View

Backlinks