The Claude Power User Field Manual

The six-month wall

Six months of daily Claude use produces a predictable disaster. Hundreds of artifacts downloaded and renamed report-final-v3 (2).md. Chats holding decisions no one can find again. Cowork sessions that edited files somewhere, the app said done, and now which version is current? The Desktop app is the center of your workday and your file system looks like a teenager's bedroom.

Prompting was never the hard part. You write good prompts; you've written hundreds. The wall you've hit is architecture: where files live, what the AI reads automatically without being asked, and which of the three Claude surfaces (Chat, Cowork, Code) should touch which work. This manual is that layer. It assumes you know what a Project is, you've saved custom instructions, and you don't need another "act as a consultant" template.

Five parts:

1. A file tree the AI can operate in
2. CLAUDE.md and where Claude reads it from
3. Claude Code for people who don't code. This is the unlock; read it even if you skip the rest.
4. Meetings: mp3/mp4 to searchable text to documents, all local
5. Hooks and skills: automation that fires without being asked

---

1. A file tree the AI can operate in

Cowork and Claude Code work on folders. Point them at chaos and you get chaos with more files in it. The fix costs one hour, once.

The root problem: humans sort files by topic, client name, department, then keep a second copy by type, all PDFs together, then give up. The AI needs neither. It needs actionability: what's live, what's recurring, what's reference, what's dead. That's Tiago Forte's PARA scheme, worth the original write-up at fortelabs.com, and it maps onto AI work better than its author could have planned:

~/work/
  projects/          <- live efforts with an end date. One folder each.
    2026-website-relaunch/
    2026-q3-fundraise/
  areas/             <- no end date: accounting/, team/, legal/
  resources/         <- reference: templates/, brand/, research/
  archive/           <- everything finished. Move it here, never delete.
  inbox/             <- everything unsorted lands here, nowhere else

Three rules make it AI-operable:

Every artifact Claude produces gets saved into the project folder it belongs to, immediately, in Markdown. Not Downloads. .md is the native format of every AI tool on earth; a folder of .md files is a database Claude can read, grep, and rewrite. Word and PDF are export formats for humans, produced at the end, not storage formats.

Dates lead filenames: 2026-07-05-board-update.md sorts itself and answers "which version is current" forever. No more final-v3 (2).

The inbox/ rule: anything unsorted goes there and only there. Once a week you empty it, or after section 3, Claude empties it for you. One dumping ground beats twelve invisible ones.

Point your Cowork projects at these folders and the "which file did it edit" problem disappears, because there's one place it could be.

2. CLAUDE.md and where Claude reads it from

Claude Code, and Cowork through it, reads files named CLAUDE.md at the start of every session, unprompted, from three levels [1]. Whatever those files say is in force before you type a word. This is the mechanism that retires your saved-prompts note:

This is automatic prompt injection, in the good sense. The instruction you used to paste at the start of every conversation now lives in a file the tool reads on its own. Deeper files add to shallower ones, so the fundraise project knows your house style and its own cap-table context without either being repeated.

What goes in the project-root one, concretely:

# ~/work/CLAUDE.md

<h2>Filing rules</h2>
- All output lands in the relevant projects/ folder as .md, filename
  prefixed YYYY-MM-DD. Never write to Desktop or Downloads.
- inbox/ is the only unsorted location. When asked to "clean inbox",
  read each file, rename to convention, move to the right folder,
  and print a table of what went where.

<h2>House style</h2>
- Short sentences. No exclamation marks. Numbers get sources.
- Vietnamese names: family name last in English docs.

<h2>When summarizing meetings</h2>
- Follow the template in resources/templates/mom-outputs.md

Notice the last line: the six-document meeting routine you've been re-pasting for months becomes a template file, referenced once, applied forever. Write these files in plain language; they're instructions, not code. And keep them short, they're read every single session, and a bloated CLAUDE.md is the most common self-inflicted wound at this level [1].

3. Claude Code for people who don't code

The name has cost more people this unlock than any UI decision in the industry. Claude Code is not a programming tool that happens to chat. It's the only Claude surface with full, direct, scriptable access to your machine: every file, every folder, every installed program. Chat can't touch your disk. Cowork runs tasks in its own managed environment. Claude Code works on your computer, as you, and your Max subscription already includes it [2].

Setup is ten minutes and none of it is programming:

# 1. Install Node.js from nodejs.org: the LTS installer, next-next-finish
# 2. In Terminal (Mac) or PowerShell (Windows):
npm install -g @anthropic-ai/claude-code

# 3. Go to your work folder and start it:
cd ~/work
claude

# 4. First run opens a browser window; sign in with your claude.ai
#    account. Max plan covers it. No API key, no card.

There's also a desktop app version with the same engine and a graphical interface if the terminal puts you off [2]. Either way, what you type into it is English:

"Read every file in inbox/, figure out which project each belongs to, rename them to the date convention, move them, and show me a table of what you did."

"Go through projects/2026-website-relaunch/ and build a decisions.md timeline out of every decision recorded in the meeting notes, with dates and who decided."

"Find every file across work/ that mentions the Hưng Yên factory and give me a one-page status brief with file paths."

That's the artifact-chaos problem from the intro, solved by the tool you'd dismissed as being for developers. It reads the CLAUDE.md files from section 2 automatically, so it already knows your filing rules when it does it.

The second superpower: Claude Code installs its own tools. It can run installers, which means it can extend itself in ways Chat never can. You say what you want; it handles the how. Which brings us to meetings.

4. Meetings: from recording to documents, all local

The recording of your Monday meeting is an .mp4 on your disk. Claude can't hear. The bridge is Whisper, OpenAI's open-source speech recognition, which runs entirely on your machine: no upload, no per-minute fees, no vendor with a copy of your board meeting. Under 8 percent word error rate, better than most paid transcription services [3].

You don't need to understand the installation. Open Claude Code and say:

"Install a local Whisper transcription tool appropriate for this machine, then transcribe recordings/2026-07-05-board.mp4 to a text file with timestamps."

It will pick the right variant for your hardware and OS, run the install, extract the audio, and produce the transcript. For reference, what it's choosing between [3]:

Then chain it, because the transcript is sitting in a folder with a CLAUDE.md that knows your meeting template:

"Transcribe every recording in inbox/recordings/, then for each one produce the six MOM outputs per the template, save them into the matching project folder, and flag any decision that contradicts an earlier one in decisions.md."

One sentence. Recording to filed documents to contradiction check, on hardware you own. The pattern generalizes: Claude Code installs a specialist tool, runs it, then does knowledge work on the output. ffmpeg for video and audio conversion, pandoc for docx and pdf, yt-dlp for pulling reference talks. You never learn these tools. You learn to ask for outcomes.

5. Hooks and skills: automation that fires without being asked

Two mechanisms move you from "Claude does what I ask" to "Claude does it whether I remember to ask or not."

Skills are folders of instructions Claude loads on demand [4]. Your meeting-processing routine, your weekly-review checklist, your deck-outline method: each becomes a named skill (/process-meeting, /weekly-review) instead of a prompt you keep in a notes app. A skill is a folder containing a SKILL.md file written in plain language. If you can write the instruction, you can write the skill; ask Claude Code to scaffold it and it will.

Hooks are commands that run automatically at fixed moments in a session, set in ~/.claude/settings.json [5]. The difference from instructions: an instruction is something Claude decides to follow; a hook always executes. Examples worth having:

- After any file edit: git-commit the change, so every AI edit is versioned and reversible. Your NextCloud sync picks it up cleanly too. - When a session ends: append what was done to a worklog.md. That file becomes your process log for academic disclosure and client billing, maintained with zero effort from you. - Before any file delete: require confirmation, always.

Ask Claude Code to write its own hooks: "Add a hook to this machine that commits every file change with a descriptive message." It edits its own settings file. The tool configures itself; that's the point of the whole stack.

The self-audit

Where you stand:

Six months of daily use typically means level 2. Everything from 3 up is an afternoon each, in order, and each level makes the next one shorter. The instructions above are the complete path. If you'd rather hand your machine to someone who's built this stack before and get it back at level 5, you know who wrote this.

---

References

1. Anthropic. "How Claude remembers your project." Claude Code documentation, code.claude.com/docs/en/memory, 2026. CLAUDE.md loading order, user/project scoping, and the guidance to keep memory files concise.
2. Anthropic. "Get started with the desktop app." Claude Code documentation, code.claude.com/docs/en/desktop-quickstart, 2026. Chat/Cowork/Code surfaces, local file access model, subscription requirements (Pro/Max/Team/Enterprise).
3. "faster-whisper vs Whisper.cpp vs WhisperX: 2026 Shootout." aifoss.dev, May 2026; OpenAI Whisper (Radford et al., "Robust Speech Recognition via Large-Scale Weak Supervision," 2022). Tool selection by hardware, word error rates, diarization support.
4. Anthropic. "Agent Skills." Claude documentation, 2026. SKILL.md structure and on-demand loading.
5. Anthropic. "Automate actions with hooks." Claude Code documentation, code.claude.com/docs/en/hooks-guide, 2026. Hook events, settings.json configuration, deterministic execution.

Tool facts verified against vendor documentation, July 2026. Forte, T., "The PARA Method," fortelabs.com, for the folder taxonomy.

Last updated: July 5, 2026