Overview
Peter Steinberger details a streamlined development routine built around Ghostty as the primary interface, selective VS Code use for lookups and occasional reviews, and Claude/Claude Code with GPT-5 for planning and code review. After experimenting with Zed, he returned to a terminal-first setup, citing preferences about Zed’s terminal presentation (link 1, link 2). He reports freezes when pasting large text into VS Code’s terminal and therefore relies on Ghostty for stability.
Hardware plays a practical role: a Dell UltraSharp U4025QW (3840×1620) display fits four Claude instances plus a Chrome window without shifting layouts, which he notes reduces window management overhead.
Tools in Daily Use
Steinberger describes a pragmatic mix of models and agents. He has been using GPT-5 to review plans and Claude for refactoring and cleanup, while decreasing reliance on Gemini due to what he calls messy editing workflows (example 1, example 2). He generally works on the main branch—after testing a worktree approach he found it slowed progress (note).
Agent counts vary by task scope: he runs 1–2 agents for straightforward work and about four agents for cleanup, testing, or UI tasks, depending on the expected “blast radius.” He remains involved in steering agents and does not see an immediate fit for background-only execution, noting that active guidance helps avoid drift.
Planning and Context Management
Two recurring practices are highlighted:
- A statusline + session ID in the terminal to recover context when switching accounts or restarting sessions (tweet, gist).
- Iterative use of plan mode, tackling smaller tasks immediately while drafting larger items for GPT-5 to review.
Prompting ranges from compact instructions to more freeform brainstorming; in the latter case, he reports that agents can produce coherent plans from loosely structured input.
Hard Parts and Custom Infrastructure
The most challenging area, according to Steinberger, remains distributed system design: dependency selection, platform choices, and defining a forward-looking database schema. To support daily work, he built custom infrastructure, including admin pages (example) and CLIs that assist both him and the agents. He reports that this investment substantially accelerated development compared to previous approaches.
Testing Strategy
Larger changes receive tests. While he considers generic automated tests uneven in quality, he finds that having the model write tests in the same context is effective at surfacing issues. The recurring point is that context is scarce and should not be diluted.
Minimal Tooling, CLI-First
Steinberger describes a “less is more” stance: he removed his last MCP after observing Claude occasionally launching Playwright unprompted when a simple code read would be faster and less context-heavy (note). He favors services with CLI tools—vercel, psql, gh, axiom—so agents can operate with minimal instructions. In practice, he keeps guidance short, e.g., a line in CLAUDE.md such as “logs: axiom or vercel cli,” and a brief database note: “psql + one example how to load env correctly so the loop is faster.” The goal is to keep the operational loop tight and avoid unnecessary tool sprawl.
Results, Comparisons, and Limits
Steinberger reports high throughput with this setup versus alternatives. He notes that Codex does not handle search as expected (reference) and says Cursor/GPT-5 review runs can take a long time without exposing intermediate reasoning, making them harder to steer. He characterizes GPT-5 as more literal in its prompting requirements—useful as a model but, in his experience, less effective as an agent without precise guidance.
He also flags upcoming rate limits effective August 28, anticipating higher costs as a result and stating no preferred alternative at present.
References and Links
- Ghostty: https://ghostty.org/
- VS Code notes: https://x.com/steipete/status/1954691261279146029/photo/1
- Zed: https://zed.dev/
- Zed terminal impressions: https://x.com/steipete/status/1955690682477134124/photo/1, https://x.com/steipete/status/1959259921439969326
- Dell UltraSharp U4025QW: https://www.dell.com/en-us/shop/dell-ultrasharp-40-curved-thunderbolt-hub-monitor-u4025qw/apd/210-bmdp/monitors-monitor-accessories
- Context/statusline gist: https://gist.github.com/steipete/8396e512171d31e934f0013e5651691e
- Work on main vs worktrees: https://x.com/steipete/status/1956340397900177686
- Gemini edit concerns: https://x.com/steipete/status/1942113964231442876, https://x.com/steipete/status/1955941943218713006/photo/1
- Admin pages example: https://x.com/steipete/status/1958956225800151465
- Codex search note: https://x.com/steipete/status/1954593439347032167
- MCP removal: https://x.com/steipete/status/1958679613489524952
TL;DR
- Terminal-first stack with Ghostty, selective VS Code usage, and Claude/GPT-5 for refactoring and plan review
- Large ultrawide monitor enables four Claude instances plus browser without window shuffling
- Iterative planning with plan mode, and statusline + session ID for context continuity
- 1–2 agents for small tasks; ~4 for cleanup, tests, or UI work
- Tests for major changes; model-authored tests in the same context catch issues
- “Less is more”: remove unnecessary tooling; rely on CLI services (vercel, psql, gh, axiom)
- Prefers interactive steering over background agents to avoid drift
- Notes rate limits taking effect on August 28; no preferred alternative identified
