Paste This Into Claude, It’ll Make You Build 10x Faster

Author: Austin Marchese | Published: 2026-06-14 | Source: transcript (15:12, 144.7K views)

Summary

Opens a genuinely new web-dev sub-theme for the wiki — AI-Assisted Development (how to drive a coding agent), as opposed to the pillar’s existing artifact-focused concepts (performance, architecture, security). Marchese frames six “power phrases” for Claude Code, leaning heavily on cited authority (Boris Cherny, Claude Code’s creator; Andrej Karpathy).

(1) “Launch sub-agents” — parallel Claude sessions, each with its own context window/prompt/permissions; use for independent perspectives (avoids one chat anchoring on its prior answer), infeasible-otherwise tasks (he ran 10 to check 10,000+ domains), and speed. (2) “Write me an implementation spec” — spec/plan before building; a spec collapses combinatorial build paths to one. (3) “Interview me” — have Claude ask the questions you didn’t know to ask, then summarize to a spec. (4) “Verify before you build” — give the agent a feedback loop and a way to see its own output (Cherny: 2–3× quality); layers = update CLAUDE.md, enable verification tools, define human-validation zones by cost of error. (5) “Build me a skill” — from validated real work, with a “gotchas” section. (6) “Automate this” — the “dangerous” one; prefer augmentation over automation (a “taste test” + 80/20 output analysis) because most automations add operational debt.

Everything here is creator-opinion/emerging — experience- and authority-based, not demonstrated — and heavily monetized (he sells BuildPartner + coaching). Credible where it echoes the cited Cherny/Karpathy/Anthropic guidance.

Key Claims

Sub-agents = parallel Claude sessions, each with isolated context/prompt/permissions. (fact)
Asking the same question repeatedly in one chat makes the model anchor and converge; parallel agents give independent perspectives. (opinion)
A detailed spec collapses combinatorics (5×5×5 → 1) so the agent gets it right. (illustrative)
Karpathy dislikes plan mode; prefers co-designing a very detailed spec (“basically the docs”) then having agents write it. (quoted)
Cherny: a feedback loop 2–3×‘s final quality; letting the agent see its output enables self-correction. (attributed)
CLAUDE.md is inserted at the start of every prompt → global behavior biasing. (fact)
Build skills from real completed work, with a “gotchas” section (per an Anthropic article). (attributed)
Prefer augmentation over automation; most automations add “operational debt.” (opinion)
Cherny calls hooks, schedules, and loops the most powerful features. (attributed)

Notable quotes

(Andrej Karpathy) “I actually don’t even like the plan mode… you have to work with your agent to design a spec that is very detailed… basically the docs, and then get the agents to write them.”

Why citable: A named authority explicitly downranking plan mode in favor of spec-driven docs — the crux of the spec-driven argument.

Connections

Entities mentioned: Austin Marchese, Claude Code, Claude, Anthropic · Boris Cherny, Andrej Karpathy, BuildPartner, Wispr Flow, Hostinger (minor) Concepts referenced: AI-Assisted Development, Clean Code & Maintainability, CI/CD & Automation (DORA Elite)

Contradictions / Tensions

No conflict with existing pages; it sits above them at the process layer. Verification/human-validation-zones relate to CI/CD & Automation (DORA Elite) quality gates; the “understand/verify” discipline echoes Clean Code & Maintainability.
Commercial bias: high — funnels to a free email series, a Claude Max giveaway for engagement, and BuildPartner/coaching. Authority-by-namedrop (Cherny/Karpathy/Anthropic) rather than independent demonstration → emerging/creator-opinion.

Notes

This is the wiki’s first source on building with AI agents (distinct from building AI features). Backfill: the cited Anthropic skill-authoring guidance and Cherny/Karpathy primary talks would strengthen AI-Assisted Development.

Paste This Into Claude, It'll Make You Build 10x Faster (6 Claude Code power phrases)