Digestable

Bite-sized principles, frameworks, and wisdom from builders, founders, and leaders I admire.

Claude Code vs Codex? I Made Them Pair-Program Instead.
Medium

Claude Code vs Codex? I Made Them Pair-Program Instead.

  1. 1.
    "Versus" is the wrong question — Don't pick a winner between Claude Code and Codex. Put both agents on the same task and let them pair-program: one writes, the other tries to tear it apart.
  2. 2.
    Cast an old workflow, don't invent a new one — Pair programming: one drives, one navigates. Claude drives, Codex navigates. The only thing that changed is a navigator who never tires, never rubber-stamps, and is there at 2am.
  3. 3.
    A generator and a critic are a pair, not competitors — Claude (Opus 4.8) reads the repo, writes in-style, and owns the shell. Codex (GPT-5) is the sharper skeptic that spots the race condition and the missed edge case. The value is the friction between them.
  4. 4.
    /duo: one command, four phases — Plan (Claude drafts, Codex critiques adversarially), Implement (Claude writes, sub-tasks fan out as reviewed patches), Adversarial review (Codex challenges the diff), and a Verify gate where the real test, typecheck, and lint must run green. No green check gets fabricated.
  5. 5.
    Log every disagreement — Each run drops a dated record of what the agents disagreed on and how it resolved. Six weeks later you can see not just what you built, but the argument that got you there.
  6. 6.
    /duox inverts the roles when your own budget is the bottleneck — Two Max plans means two token budgets. When Claude is scarce, Codex does the heavy lifting and Claude just orchestrates. Same discipline, opposite economics: burn the side of the ledger with room.
  7. 7.
    The disagreement is the actual product — Agreement is cheap. When the driver and skeptic clash, the flag is either real (fix it), wrong (Claude explains in writing), or needs a human, and it lands on your desk already framed.
  8. 8.
    Match the tool to the risk — Low-risk change: a single agent, no ceremony. Money, auth, data integrity, migrations: the full pair, every time. Strategy or advice: neither, just answer. Budget-constrained heavy work: the inversion.
  9. 9.
    The leverage is the system, not the model — A clear plan, an adversarial review you can't skip, a verify gate that doesn't lie, and a record of every argument. The models keep getting better; the discipline is what makes them trustworthy on work that matters.
Joshua Bascos
AI-Assisted DevelopmentClaude CodeCodexPair ProgrammingSoftware Engineering
a16z and founderspodcast

Elon Musk's Company-Building Principles

  1. 1.
    Apply The Algorithm constantly — a. Question every requirement. b. Delete any part of the process you can. c. Simplify and optimize. d. Accelerate cycle time. e. Automate.
  2. 2.
    The only rules are the ones dictated by the laws of physics — Everything else is a recommendation.
  3. 3.
    Compare the cost of a finished product to the cost of its raw materials — The Idiot Index. If a product has a high index, its cost could be reduced through more efficient manufacturing techniques.
  4. 4.
    If a timeline is long, it's wrong
  5. 5.
    Hire for attitude — Skills can be taught.
  6. 6.
    The leader should be on the front lines
  7. 7.
    Camaraderie is dangerous — It makes it hard for people to challenge each other's work.
  8. 8.
    Don't fear losing — It hurts the first 50 times, but then you'll be able to play with less emotion and take more risks.
  9. 9.
    Design and production should never be separated — Keep everything together and feedback immediate.
  10. 10.
    Stay heads-down focused on doing useful things for civilization
  11. 11.
    The mission comes first — Keep the entire company committed to a common goal.
Elon Musk, CEO of SpaceX
LeadershipFirst PrinciplesStartupsManufacturingManagement