# Files Over Prompts
Build HTML Artifacts and LLM Wikis Your Claude Agent Actually Reads

- Price: $12.99 (Kindle ebook)
- Level: Intermediate
- Pages: 196
- Best for: Prompt writers whose context dies with the session
- You finish with: A 5-entry LLM wiki, persistent memory, and three throwaway HTML editors in one weekend
- Series: AI Memory Playbooks, book 3 of 4
- Buy on Amazon: https://amzn.to/4v1TkDs
- Web page: https://youcanbuildthings.com/books/files-over-prompts

## What you get

Prompts plateaued. Files compound. Build the substrate the smart developers already shipped: a 5-entry LLM wiki by Saturday morning, persistent memory by Saturday night, three throwaway HTML editors by Sunday. No SaaS, no API except Anthropic, every file under your control.

Prompts plateaued. Files compound. The first complete playbook for the era after prompts: a 5-entry LLM wiki by Saturday morning, persistent memory across every Claude session by Saturday night, and three production-grade HTML editors by Sunday. 196 pages, every chapter shipping a working artifact, every file on disk under your control.

## Chapters

1. The Inversion: Why you're scrolling past Karpathy's wiki wrong, and what changes when the agent reads first.
2. Foundation: Writer-mode vs reader-mode and the 4 rules for organizing data for the reader who isn't you.
3. Build Your First LLM Wiki: The Karpathy skeleton on disk in 90 minutes: 2 folders, a schema file, and 5 entries across PKM, sales, support, code review, and research.
4. Persistent Memory: Three injection mechanisms (file import, Notion + iOS Shortcuts, MCP server) and the Tuesday morning test that proves it works.
5. Retrieval, Linting & Hallucination Compounding: How one wrong summary becomes 16 downstream errors and the weekly lint pass that cuts the tree at level 1.
6. HTML Artifacts, The Pivot: Why you'll stop writing markdown: the 6 reasons, the 6 FAQ counter-objections, and the design-system file every later artifact pulls from.
7. HTML Specs, Plans & Reports: Implementation plan, weekly status report, and research explainer as HTML artifacts your team actually reads.
8. HTML Code Review & Understanding: The PR review HTML you attach to every PR, plus the code-tour HTML for unfamiliar modules.
9. HTML Design Artifacts: Component variants, animation prototypes, SVG flowcharts, and slide decks: the four design-adjacent categories where markdown loses.
10. Custom Editing Interfaces: Three throwaway editors: triage board, feature flag editor, prompt tuner. Each ends with an export, each disposable in the best sense.
11. The Web of HTML Files: Multi-agent, multi-vault, team scale: the verification agent, three-vault topology, and the no-code path for non-developers.
12. When Personal AI OS Is Overkill: The cost ladder, the decision tree, and the maintenance routine that keeps the substrate honest past week four.

## Free guides from this book

- [Build a Throwaway Claude Prompt Tuner in One Prompt](https://youcanbuildthings.com/articles/claude-prompt-tuner-html-editor.md): Build a claude prompt tuner as one HTML file: template with slot highlighting, three live-rendering examples, token counter. Thariq's verbatim prompt + fix.
- [How to Build Karpathy's LLM Wiki in 90 Minutes](https://youcanbuildthings.com/articles/karpathy-llm-wiki.md): The karpathy llm wiki is two folders and a schema file, not three folders. Build the real structure on disk in 90 minutes with the mkdir commands and 5 entries.
- [LLM Wiki Maintenance: One Wrong Summary, 16 Errors](https://youcanbuildthings.com/articles/llm-wiki-maintenance-lint.md): Llm wiki maintenance fails when one bad summary cascades. The 1 to 4 to 16 error pattern, the lint pass that cuts the tree, and the cost math at every scale.
- [Persistent Memory for Claude Code in 30 Minutes](https://youcanbuildthings.com/articles/claude-code-persistent-memory.md): Wire claude code persistent memory in 30 minutes. The memory-pack.md pattern, the four CLAUDE.md scopes, and the Tuesday morning test that proves it loaded.
- [Ship HTML Artifacts on Every PR (Claude Code Pattern)](https://youcanbuildthings.com/articles/html-artifacts-claude-pr-review.md): Html artifacts claude engineers attach to every PR. The verbatim Thariq prompt, three first-run failure modes, and the 20% surfaced first that beats markdown.