Codex vs Claude, $50k Consulting, Hermes SDK Fix — AI Daily May 26
565 messages · 90 active members
Overview
Topics
Power users reported fully switching to Codex for stubborn bugs and tight instruction sets, while Claude defenders held ground on fire-and-forget multi-step orchestration. @geilt detailed how Codex syncs CLI/app/phone sessions, pairs with worktrees, and shares SSH connections. @justingacina explained using AGENTS.md (now read by Codex natively) as an entry point to progressive-disclosure context directories rather than dumping everything inline.
Members shared real numbers: $50k for 2-day insurance training, $1M for a 4-month build, $25k/day Wall Street courses. Key insight: under-pricing kills credibility, and abrasive pitches ('any Indian dev can build it for $100, you're paying for knowing what to build') close deals. Biggest projects come from billionaire desires, not typical SaaS budgets.
@Wootbro shared a /goal OMO Ultrawork + TEAM MODE template covering preflight, planning, 5-10 parallel agent swarms with collision control, fan-in verification, and closeout audits. Separately, Hermes users hit 'NoneType is not iterable' errors with GPT-5.5 because Hermes pins openai==2.24.0 — community fix: pip install openai==2.38.0 in the venv, restart gateway and TUI.
@tylertrowbridge sparked comparisons between Mac Studios, Hetzner (recently raised prices), OVH ($10/16GB), and Contabo. @Tz1888 recommended Mosh + Tmux via Termius for persistent sessions accessible from laptop and phone, with zram tricks suggested for doubling effective RAM on tight-budget boxes.
Workflow shared: 10-15 GPT Image 2 storyboards → Seedance 1.5 Pro (cheaper) or Kling 3 start/end frame animation, fully scriptable via Codex/Claude. Veo 3 voice consistency remains painful. Parallel debate sparked by an OpenAI vs Anthropic profitability article: single-feature SaaS gets one-shot by foundation model updates, so skills stacking and client relationships beat niche-tool bets.
Key Takeaways
- Codex is winning power users for stubborn bugs and instruction-heavy work; Claude still leads for autonomous multi-step orchestration. Use AGENTS.md — most platforms now fall back to it.
- Enterprise AI consulting clears $25k-$50k/day for basic dashboards and automations — credibility scales with price, not complexity. Under-pricing kills deals.
- If your Hermes broke on GPT-5.5 today: pip install openai==2.38.0 in the venv, restart gateway and TUI. Rollback path documented.
- Exploit forced-context files (claude.md, agents.md) as entry points to a progressive-disclosure directory; pair with Mosh+Tmux+Termius on a $10-20/mo VPS for nomad-friendly persistent sessions.
- Single-feature AI SaaS is high-risk against foundation model updates — stack skills and client relationships that compound across model generations.
Hot Threads
$50k for 2 days teaching basic AI to insurance execs — how to price and pitch enterprise consulting
Mac mini vs VPS for nomading AI builders
GPT-5.5 + Hermes SDK breakage — community-debugged patch and rollback