How to Build a Claude Skill From Scratch — Write With AI

Why this is in the vault

Direct hit on RDCO’s core practice: skill engineering. Write With AI is a K-whitelist sender on writing craft, and this issue collapses the writing/skill-building distinction in a way that maps cleanly onto our ~/.claude/skills/ discipline. The promo for their “Low-Ticket Launchpad” bootcamp is heavy at the bottom (PS section is mostly a sales CTA), but the top 70% is real craft material — not a sales-funnel wrapper. Worth filing for the central thesis even if the worked example (Cowork + Notion) is downstream of where we already operate.

The core argument

“Great prompts are education products copy-pasted into AI.”

Bush’s claim: when he writes a 5,000-word course module teaching writers how to do X, the same module — wrapped with a one-line invocation at the top — becomes a 5,000-word prompt that produces excellent output. Same artifact, audience swapped from human to model.

Therefore:

• Writing well = using AI well. The skill of explaining clearly to a human IS the skill of articulating to AI.
• Books, courses, digital products are long-form prompts for humans. A book changes behavior; a prompt chain does the same — just for a robot reader.
• Don’t let AI take the wheel. When Claude says “I’d recommend X,” answering “sounds good” produces mediocre skills. The author keeps the wheel; AI executes.
• Build skills one step at a time. Most people build the whole skill at once, get janky output, spend three days debugging. Instead: build step 1, test with a real example, ship to Notion if good. If bad, the problem is your documentation — fix the input, retest. Then move to step 2.

His worked example (omittable for our purposes): a Notion + Claude Cowork “Digital Product AI Hub” with four steps — generate ideas, run Offer Creation Checklist, build outline via Bookends Framework, write modules via 7-Section Module Template.

The closing claim: “The writers who figure this out are about to have a massive advantage. Not because they’re better at AI. Because they’re better at thinking.”

Mapping against Ray Data Co

Strong mapping. This article is the writer-side restatement of 2026-04-11-garry-tan-thin-harness-fat-skills. Tan: skills are markdown documents that teach the model a process. Bush: skills are course modules wrapped with a one-line invocation. Same artifact, different vocabulary. Both converge on: the value is in the document, not the harness.

Specific RDCO implications:

1. Validates the “skills as written documents” architecture. Every skill in ~/.claude/skills/ is, by Bush’s framing, a long-form teaching document for a model. The process-newsletter skill executing this very task is a prose explainer of how to triage, classify, file, and cross-link — written for me to read like a smart-friend-over-coffee briefing. Bush’s framing is correct: when our SKILL.md files read like teaching, they perform; when they read like terse config, they don’t.

2. Reinforces the “build one step, test, ship” loop. Matches our /skillify skill’s pattern of extracting workflow from a just-completed conversation rather than designing the whole skill abstractly upfront. Also matches Klaassen’s compound-engineering “fix the input, not the model” framing.

3. “Don’t let AI take the wheel” maps to founder-as-advisor-not-pair-programmer. When the founder writes Sanity Check pieces or RDCO copy, the skill is to keep the wheel. When I’m building or modifying skills, the same applies — the founder’s judgment is load-bearing on direction, not on syntax.

4. The writing/thinking equivalence is the operational thesis behind the vault itself. The vault is where the founder does the hard thinking once, in prose, so it can be referenced later — by him and by me. Bush’s “write more” prescription is what /process-newsletter enforces structurally: distill, write the assessment, file it, cross-link. The act of writing the note IS the act of integrating the source.

5. Cowork+Notion vs ~/.claude/skills/. Bush’s stack (Cowork + Notion as the chained-prompt hub) is one rung below where we sit. We use SKILL.md files in a real harness with hooks, MCP servers, and durable state — Tan’s “thin harness, fat skills” architecture. Bush is teaching writers how to climb to where we already are. Their stack is the on-ramp; we’re the destination. No tooling change indicated.

Net: useful as a reinforcing voice and as a writer-friendly translation we can point external collaborators to. Not architectural news.

Sponsor / self-CTA detection

• Self-CTA, heavy: PS section pitches “Low-Ticket Launchpad LIVE” bootcamp ($-tier inferred) starting “tomorrow” (i.e., 2026-04-27). Five Substack-redirect links to their own properties (Write With AI, Ship 30 For 30, Premium Ghostwriting Academy, Typeshare, Ghostbase) in the signature.
• No third-party sponsor. All commercial intent is author-owned product.
• Treat as standard Bush/Schafer cadence: educational top, bootcamp pitch bottom. The top 70% stands alone.

Related

• 2026-04-11-garry-tan-thin-harness-fat-skills — the architectural source-of-truth this article echoes in writer-vocabulary
• 2026-04-22-garry-tan-skillify-it-workflow — the “extract a skill from a successful workflow” pattern; complements Bush’s “build one step at a time” loop
• commentary-tan-fat-skills-thin-harness-2026-04-14 — RDCO’s own commentary on the Tan thesis
• 2026-01-07-write-with-ai-headline-claude-skill — earlier Write With AI piece on a specific Claude skill (headline writing)
• 2026-02-15-write-with-ai-cowork-bootcamp-promo — prior Bush bootcamp promo for context on cadence/voice
• 2026-01-13-every-claude-cowork-review — Cowork primer (the tool Bush builds on)