Explore what to build next

Why do AI-generated interfaces always have that "flavor"? Purple-blue gradients, rounded cards, emoji everywhere. Taste isn't talent — it's rules that can be written down: lead with references, design tokens first, typography decides 80% of the look. AI has no taste, but it can execute yours.

On October 1, Anthropic announced an expanded partnership with Barclays: half of the bank's developers will use Claude Code by year-end, most by end of 2027. When even heavily regulated banks bet core productivity on AI coding tools, "AI writes code" has officially entered the scaled-deployment phase.

AI now writes in a day what a team wrote in a week — but "written" doesn't mean "owned." This guide covers code reading for the AI era: get the map first, start from entry points and tests, keep a "distrust checklist." Reading code isn't nitpicking; it's being able to catch AI when it's wrong.

On October 1, Microsoft Azure Research released 301,026 real Copilot coding-agent session traces — 9.3M model calls, 8.7M tool calls, fully anonymized. AI coding finally has production-grade workload data, and the bills for caching, retries, and idle scheduling can finally be calculated.

"Solo company" is 2026's sexiest phrase, but few do the other math: true costs. Subscriptions at $200–$800/month are the small part; the real cost is "time tax" — support tax 1–2h daily, ops tax 3–5h weekly, finance & tax 1–2 days monthly, compliance & legal per incident, emotional tax 24/7. Total: 15–20 hours a week not building product. Plus three tax-reduction strategies: make it visible, accept "solo ≠ alone," budget for emotions.

On September 30, Copilot CLI, Claude Code, Antigravity CLI, and Kiro CLI all shipped — plus Gemini CLI and Codex CLI's v0.159 series the day before. Five terminal agents updating within 48 hours isn't coincidence; it's proof "terminal-first" became industry consensus. This piece rounds up the key releases (Codex 0.159's keyword: "control") and asks why every major lab is pouring resources into the terminal in H2 2026.

Context windows keep growing in 2026, but "stuff everything in" buys sky-high bills and diluted attention. This hands-on guide treats context engineering as managing three ledgers: token money (cache everything cacheable, retrieve instead of stuffing, compress history), model attention (important at both ends, structured expression, context decay, reference over copy), and your time (10-minute cap for one-offs, build assets for recurring tasks). Includes a 5-minute context checklist template.

DevDay's most underestimated launch wasn't a model — it was a login button: "Sign in with ChatGPT," letting Plus/Pro subscribers spend their allowance inside 16 partner tools. On September 29 Windsurf followed, wiring it into Devin Desktop. This piece analyzes OpenAI's "cash-register strategy": trading one subscription's convenience for settlement rights over the whole AI tool ecosystem. History rhymes — the App Store played this game.

Vibe coding's sweetest moment is the demo — but sign a toB contract and start delivering, and a chasm opens between demo and delivery. This piece draws four boundaries: security compliance is a hard gate (quote at 5x demo cost), maintenance liability is bottomless (never promise a free year), requirement changes must be billed (block the "quick change"), data ownership and IP go in the contract. The thesis: between "demo" and "delivery" lies a SOC 2 — and SOC 2 can't be vibed.

Buried in the Claude Code v2.1.277 changelog from September 18: AGENTS.md support — projects without CLAUDE.md now get AGENTS.md read instead. It looks like "one more filename recognized"; it's actually an industry signal: the standards war over agent instruction files is converging, likely on OpenAI's AGENTS.md. This piece explains why it matters and what a good AGENTS.md looks like.

The coding workflow changed: the agent writes, you review. But reviewing AI code line-by-line no longer works — AI rarely writes syntax errors; its specialty is "looks done, actually isn't." This checklist gives 12 checkpoints: 3 for requirement alignment, 4 correctness traps, 3 engineering quality, 2 security red lines. The thesis: reviewing AI code isn't about finding bugs — it's about finding "things you thought it did, but it didn't."

In the second week of September 2026, eleven disclosed US venture rounds totaled roughly $10.8B, with AI coding taking the highlights: Cognition $2B ($48B valuation), Factory $200M ($5B, 3x in five months), Temporal $550M ($12.55B). This piece skips the "how much" noise to ask what the money is actually betting on: capital is flowing from "personal coding assistants" into "enterprise software factories" — with direct consequences for everyone who codes.