Thanks for commenting LEVEL. Here are the files from the post: a shared memory template, 3 starter skills, a ban list, and my subagent and schedule prompts.
CLAUDE.md and keep it short. In Claude Code, put it in your project folder or in ~/.claude/CLAUDE.md. In the Claude app, paste it into a Project's instructions.~/.claude/skills/<name>/SKILL.md (Claude Code), or zip the folder and upload it under Settings → Capabilities → Skills (Claude app; menu names can differ).CLAUDE.md or a Style.CLAUDE.md and the ban list are instructions, not locks. Claude follows them most of the time.
# CLAUDE.md: about me and how to work with me
<!--
HOW TO USE THIS TEMPLATE
- Fill in each [bracket], then delete what you don't need.
- Claude Code reads CLAUDE.md at the start of every session. Put it in your project folder (that project only)
or in ~/.claude/CLAUDE.md (all your projects).
- In the Claude app, paste the filled-in sections into a Project's instructions instead.
- Keep it short: well under 200 lines. Long files get followed less well.
- Corrected Claude on the same thing twice? Add one line here.
- Pasting the same steps again and again? Make it a skill (SKILL.md), not a longer CLAUDE.md.
- This file is a set of instructions, not a lock. A "never" rule written here is a strong request.
If something must NEVER happen (for example, editing a certain folder), enforce it with a hook
in Claude Code. A hook is a script that runs before or after an action and can block it.
-->
## Who I am
- I'm [name], a [role, e.g. creator / founder / developer] working on [what you make].
- My audience is [who they are] who want [what they want].
- Right now I'm focused on [current goal or project].
## My voice
- Tone: [e.g. short, direct, friendly, no hype].
- Sentences: [e.g. short; one idea per sentence].
- I sound like: [paste 2-3 lines you actually wrote, as examples].
- Never use the words in my ban list: [paste it here, or link ban-words.md].
## Rules
- Ask me before [big edits / deleting files / anything you can't undo].
- If something is unclear, ask one short question instead of guessing.
- Don't invent facts, numbers, quotes or sources. Say when you're not sure.
- [Add one line each time you correct the same thing twice.]
## How to work
- For anything bigger than a small fix: make a short plan first and wait for my OK.
- Give one job to each helper (subagent). Don't run many at once unless I ask.
- Before you say "done," show proof: what you checked and how (test, preview, re-read against these rules).
- Use my skills when they fit: [list your skills, e.g. /brief-builder, /repurpose-content, /proof-check].
## Project notes
- Key files and folders: [e.g. /drafts, /brand, /scripts].
- Tools and accounts in use: [e.g. Notion, Google Drive, GitHub].
- Done looks like: [e.g. a draft in /drafts, checked with /proof-check].
---
name: brief-builder
description: Turns a messy idea, voice note transcript or brain-dump into a clear one-page brief and a short plan before any work starts. Use when the user shares a rough idea for a post, video, product, feature or project, or says "help me plan this."
---
# Brief builder
Goal: get from "messy idea" to "agreed brief" before anything gets made.
## Steps
1. Read everything the user gave you. Don't start making the thing yet.
2. Pull out what's already clear: the goal, the audience, the format, the deadline, and any must-haves.
3. List what's missing or unclear. Ask at most 3 short questions, the most important first. If the user says "just guess," make sensible guesses and mark them as `(assumed)`.
4. Write the brief in this format:
```
## Brief: <short title>
- Goal: <one line: what success looks like>
- Audience: <who it's for and what they want>
- Format: <post / carousel / video / feature / doc, plus length or size>
- Key message: <the one thing they should remember>
- Must include: <bullets>
- Avoid: <bullets>
- Deadline: <date or "none">
- Done when: <how we'll check it's finished>
```
5. Add a plan of 3-7 numbered steps, each with what it produces.
6. Stop and ask: "Does this brief look right? I'll start when you say go." Don't start the work until the user approves.
## Rules
- Keep the brief to one screen.
- Use the user's own words for the goal and key message where you can.
- Never invent facts, numbers or quotes to fill a gap. Mark the gap instead.
---
name: repurpose-content
description: Turns one finished piece (an article, script, transcript, thread or long post) into several platform-ready drafts in the user's voice, such as a carousel outline, a short-video script, a thread and a newsletter blurb. Use when the user asks to repurpose, cut down or spread one piece of content across platforms.
---
# Repurpose content
Goal: one strong piece in, several native drafts out, all in the user's voice.
## Steps
1. Ask which platforms they want if they didn't say. Default: carousel, short video, thread, newsletter blurb.
2. Read the source piece. Pull out:
- the main idea in one sentence
- 3-5 key points
- the best line or example (the "hook material")
- any facts or numbers, each with its source
3. Write one hook per platform. Each hook should state a specific pain, result or surprise, not a vague topic.
4. Draft each format:
- **Carousel:** a cover headline (max ~8 words), 5-8 slides with 1 short headline and up to 3 short points each, then a CTA slide.
- **Short video (30-60 s):** numbered beats with on-screen text plus what's said; hook in the first 2 seconds.
- **Thread:** a hook post, then 4-8 posts with one point each, then a CTA.
- **Newsletter blurb:** 80-120 words and one link-style CTA.
5. Apply the user's voice and ban list from CLAUDE.md (or the ban-words file), if present.
6. Run a quick self-check before handing over: every fact traces back to the source, nothing is invented, each hook is under ~12 words, and no banned words appear.
7. Deliver the drafts under clear headings, followed by a 1-line note on anything you weren't sure about.
## Rules
- Rewrite for each platform; don't paste the same text everywhere.
- Never add stats, quotes or claims that aren't in the source. Label opinions as opinions.
- Keep on-image text short. Image generators misspell long text.
---
name: proof-check
description: Checks finished work against the request, the user's rules in CLAUDE.md and the ban list, then reports proof of what was checked before anything is called done. Use before delivering any draft, code change, file or plan, or when the user asks "is this ready?" or "check your work."
---
# Proof check
Goal: never say "done" without proof.
## Steps
1. Restate the request in one line, along with its "done when" (from the brief if there is one).
2. Check the work against each point and mark it pass, fail or unsure:
- **Request:** everything asked for is there, and nothing extra was added that wasn't asked for.
- **Rules:** every rule in CLAUDE.md is followed (voice, length, format, "ask before" rules).
- **Ban list:** search the text for each banned word and phrase. List any hits and fix them.
- **Facts:** every number, quote, name and feature claim has a source or is labelled unverified.
- **For code:** run the tests or the build if they exist, and report the command and result. If you can't run them, say so.
- **For files:** confirm each file exists at its path and opens.
- **For text on images:** check the spelling word by word.
3. Fix every fail that you can fix. Re-check those points.
4. Report in this format:
```
Proof check: <PASS / PASS WITH NOTES / NOT READY>
- Checked: <what you checked and how>
- Fixed: <what you changed>
- Open: <anything unsure or unfixable, and what's needed>
```
## Rules
- Never mark something pass that you didn't actually check.
- "Unsure" is fine. A fake "pass" is not.
- Keep the report short; details go under Open only if they matter.
# Ban-word list: cut the AI slop
## How to use it
- **Claude Code:** paste the list (or the parts you care about) into your CLAUDE.md under "My voice". You can also put it in an output style, Claude Code's version of a style. Only one output style is active at a time.
- **Claude app:** add it to a Project's instructions, or create a custom Style that includes the list.
- **It's a request, not a filter.** Claude will mostly follow it but can slip. Run the `proof-check` skill before you publish.
- **Advanced:** in Claude Code, a hook can scan files Claude writes and flag banned words. A hook can't rewrite Claude's chat replies.
- Trim the list to your own pet peeves. A shorter list gets followed better than a huge one.
## Words to cut (with plain swaps)
| Instead of | Try |
|---|---|
| delve / dive into / deep dive | look at, cover, break down |
| unlock / unleash | get, open up, start |
| elevate / supercharge | improve, speed up |
| leverage / harness / utilize | use |
| seamless / seamlessly | smooth, easy (or cut) |
| robust | solid, strong (or cut) |
| game-changer / revolutionize | name the actual change |
| cutting-edge / state-of-the-art | new, latest (or cut) |
| landscape / realm / sphere | field, world (or cut) |
| tapestry / symphony / testament to | cut; say the plain thing |
| embark / journey | start, process |
| pivotal / crucial / paramount | key, important (or cut) |
| foster / cultivate | build, grow |
| intricate / multifaceted / nuanced | detailed, complex (or explain how) |
| meticulous / meticulously | careful (or show the care) |
| vibrant / bustling | cut, or describe what's there |
| boost / skyrocket | grow by <number> (if you have it) |
| myriad / plethora | many, lots |
| furthermore / moreover / additionally | also, and (or start a new sentence) |
| ultimately / essentially | cut |
## Phrases to cut
- "In today's fast-paced world…" / "In the ever-evolving world of…"
- "It's important to note that…" / "It's worth noting…"
- "Let's dive in." / "Buckle up." / "Here's the thing:"
- "Great question!" / "Certainly!" / "Absolutely!" at the start of a reply
- "I hope this helps!" / "Feel free to reach out…"
- "Whether you're a beginner or a pro…"
- "Not just X, but Y." (when it's used as filler)
- "In conclusion…" / "In summary…" at the end of short pieces
- "Unlock the power of…" / "Take your <thing> to the next level"
## Patterns to cut
- Lists of three when one or two words would do ("fast, simple, and powerful").
- Sentences that open with a question you then answer right away ("The result? …"), when it's overused.
- Emoji at the start of every heading.
- Long dashes everywhere. Use a period or a comma.
- Ending with a moral or summary the reader already got.
Use it in Claude Code when a job has parts that do not depend on each other.
Split this job into independent parts. Run one subagent per part, in parallel. Give each subagent: the goal, the files to read, what is already ruled out, and what a finished result looks like. When they come back, check each result yourself before you merge anything. Tell me which parts you did not verify.
Use it in any Claude surface that has scheduled tasks (the name and the menu differ by app and plan).
Every weekday at [8:00], do this without asking me: [the task]. Read CLAUDE.md first. Write the result to [file or place]. If something fails or you are not sure, stop and tell me what happened instead of guessing. Keep the report under 10 lines.
Follow @mhtua for more AI systems and workflows.
Follow @mhtua on Instagram. Comment the keyword on any post and I'll send you the pack.
FOLLOW @MHTUA