NateWiki — Open Brain corpus
Home · The Doctrine · All topics · Search
priority 1 31 posts

Claude Code & Coding-Agent Practice

The playbook

This cluster runs from February 2025 (the Cursor/Windsurf/Lovable/Bolt era) through July 2026 (harness audits and token-saver skills), and Nate's position visibly moves through three distinct eras rather than holding one static view.

Era 1 — vibe coding as a prompting skill (Feb–Aug 2025). Early posts treat AI-assisted coding as a tool-selection and prompting problem: pick Windsurf/Cursor/Lovable/Bolt/Replit for the right job, keep prompts short and iterative rather than monolithic, chain prompts (skeleton → auth → database → optimization) instead of asking for everything at once, and always feed the AI the actual error text rather than "it doesn't work, fix it." The recurring warning in this era is that AI is a confident junior dev, not an architect — it hallucinates functions, omits security/validation unless told, and needs git checkpoints before every change. Claude Code's June 2025 launch reframes the whole category: Nate argues it's a general-purpose agent wearing a coding costume, not a coding tool — the real unlock is planning-first workflows ("think hard," explicit plan-then-execute), CLAUDE.md as persistent memory, and treating the terminal as a conversational interface non-engineers can use too.

Era 2 — reading code and engineering discipline overtake writing code (May–Oct 2025). As generation got cheap, Nate's emphasis flips: "Taming the Genie" and "Code as Story" argue that reading/tracing architecture (follow one datum, map boundaries, interrogate hand-offs) is now the durable skill, since AI floods repos with code nobody vets. "Software Engineering Isn't Dead" pushes back on both AI-doomer and AI-utopian takes: engineers must own invariants, production observability, and regulatory-explainable failure — AI raises the bar on engineering rigor rather than removing it. The "14 AI Questions" post explicitly reframes "which tool" as the wrong question — team discipline (design docs, review rigor, security review of AI code) determines outcomes far more than model choice.

Era 3 — architecture over model, harness over prompt (Oct 2025–Jul 2026). The Claude-vs-Codex framing recurs three times (Oct 2025, Feb 2026, Apr 2026) and hardens into a stable axis: Claude Code/Cowork = iterative, collaborative, coordinate-across-tools, cheap-to-course-correct; Codex = delegate-and-walk-away, isolated-sandbox, correctness-optimized, better for independent/parallelizable/GUI-driving work. Route by task shape, not brand loyalty. Underneath that, the "harness beats model" thesis (same Opus scoring 78% vs. 42% depending on scaffold) becomes central, alongside a skills/plugins maturity ladder (prompt → skill → plugin → MCP connector → deterministic script) and a hard correction in mid-2026: harnesses that accumulate one correction at a time become bloated and start failing delivery contracts even as they reason better — the fix is auditing what's actually in your CLAUDE.md/skills, deduplicating rules, and turning binary requirements into enforced checks (validators, hooks) instead of more prose. Parallel threads: token/context hygiene (fresh conversations, markdown-not-raw-PDF, /context audits, model routing Opus/Sonnet/Haiku) as real budget levers; tool/API latency (not model speed) as the Amdahl's-law bottleneck once inference gets fast; "dark code" (AI-shipped code nobody ever comprehended) as a named failure mode needing spec-driven development and PR-time comprehension gates; and a trust inversion (Mozilla's Mythos finding 271 Firefox bugs) suggesting adversarially-verified AI code may become more trustworthy than unreviewed human code, making codebase legibility a security deadline.

For an owner-operator running an agent fleet, the synthesis is: don't pick a permanent "best" tool — route Claude Code/Cowork to multi-tool, interdependent, judgment-heavy work and Codex-style delegation to bounded, parallelizable, or GUI-only tasks; invest in the harness (CLAUDE.md, skills, MCP connectors) as seriously as the model choice, but audit and prune it on a schedule rather than only ever adding rules; convert your best repeated judgment calls into skills, your cross-system workflows into MCP connectors, and your genuinely deterministic steps into scripts/hooks rather than standing prompts; treat token spend and context bloat as a hygiene problem you actively manage, not a tax you absorb; and budget real review time for AI output specifically at the seams (auth, data boundaries, error handling) where hallucinated-but-plausible code causes the most damage.

Key moves

The posts

Related