TIP 01 · CONTEXTCLAUDE.md — the file that stops you from repeating yourself to the AI

Bottom line. Once I dropped a CLAUDE.md into the folder, I stopped saying the same thing twice. Context started carrying over like a bridge between sessions.

When I first started running Claude Code, I introduced myself again every single day. A new session wasn’t an old friend — it was a complete stranger. Every rule and preference I’d agreed on yesterday was just gone.

Then I noticed something. If I dropped a CLAUDE.md file in the folder, every new session would automatically read the conventions I’d written there. I found the pattern documented in the official CLAUDE.md memory docs.

~/Documents/Claude/Projects/
├── automation/CLAUDE.md           # Main ops rules
├── real-estate-blog/CLAUDE.md    # RE tone & tax source rules
├── ediblog/CLAUDE.md             # 1st-person 80% gate
└── jessinvestment/CLAUDE.md      # EN mirror translation rules

After running it for a while, my CLAUDE.md naturally settled into 5 sections: Identity / Hard Rules / Soft Rules / Commands / Recent Changes.

# CLAUDE.md

## Identity
This folder runs the "[blog name]" blog / Operator: Eddie, solo

## Hard Rules (non-negotiable)
- No auto-publish (drafts only)
- API keys only loaded from secrets/ folder

## Soft Rules (preferred)
- 1st person 80%+ tone
- At least one official external source cited

## Commands (master index)
| Command    | Description                       |
| /run-daily | Generate daily publish queue      |

## Recent Changes
- 2026-05-12: Skills split out; CLAUDE.md slimmed down

I used to try to make it perfect from the start. That actually held me back. Once I started cycling result → review → refine a lot, CLAUDE.md grew on its own without me forcing it.

One trap I hit, though: when CLAUDE.md got too big, every new session loaded the whole thing and token cost kept piling up. That pushed me to adopt the Skills offload from TIP 03 as the next step.

→ The full story of how my CLAUDE.md grew from 10 lines to 300 is coming in Part 2.

TIP 02 · SHORTCUTSlash commands — if you’ve pasted the same prompt 3 times, extract it now

Bottom line. The moment I find myself typing the same prompt for the third time, I pull it into .claude/commands/ right there. That’s the essence of slash commands.

Before slash commands, I pasted the same prompt every single time. Time was one cost, but the bigger one was skipping one line by accident and getting a completely different result.

Now I just type /run-daily. I learned how to build them from the official docs on slash commands.

---
description: Generate the daily publish-candidate queue and send a review request to Telegram
---

Follow this procedure:

1. Scan the drafts/ folder and count unpublished posts
2. If under 5, write a new post (trigger the blog-style-guide skill)
3. Send a preview of the new post to Telegram
4. On user approval, save it as a draft

Typing /run-daily kicks off exactly that. This is the actual command skeleton I use. I found that the one-line description is what decides whether the command gets triggered. The description field in the official frontmatter spec doubles as the trigger keyword:

$ /run-daily
[Claude] Scanning drafts/ folder...
Found 3 unpublished posts. Queue under 5 → writing 1 new post.
Triggering blog-style-guide skill...
✅ Generated AI_update_20260515.html (8,234 chars)
Telegram preview sent ✓
Review requested — waiting up to 30 min for reply

→ My full playbook for running 30 commands is coming in Part 3.

TIP 03 · OFFLOADSkills — a heavy CLAUDE.md blows up your token bill

Bottom line. What I need every session goes into CLAUDE.md. What I need occasionally goes into Skills. That’s where my token diet started.

I knew CLAUDE.md was powerful, but I was slow to notice one trap: it gets loaded in full at the start of every session, burning tokens each time. The more ops rules I added, the more the cost compounded.

So now anything I don’t need on every turn lives in a Skill. Claude pulls it in only when it’s actually needed. The official Skills docs have the full pattern.

Right now I run 3 Skills to keep CLAUDE.md slim.

---
name: add-launch-job
description: Checklist for adding, editing, or removing launchd jobs (LaunchAgents).
  Procedures for new blog/session setup, one-shot job cleanup, plist creation/deletion.
---

# launchd job ops checklist
...

Here’s the actual classification table I use. It follows the progressive disclosure pattern from the official Skills guide:

Info type                          Goes in
─────────────────────────────────────────────────────
Ops rules · 1st-person tone guide   CLAUDE.md (every session)
Master command index               CLAUDE.md (every session)
New-job checklist                  Skills (occasional)
Incident · captcha recovery        Skills (only when needed)
Blog writing style guide           Skills (only when writing)
v2 infographic component rules     Skills (only when designing)

Here’s an auto-trigger in action. I rely on description-keyword matching so the skill wakes up on its own:

User: Add a new launchd job
Claude: [add-launch-job skill auto-triggered]
        Applying checklist...
        1. Write .job file (TIMES/COMMAND/WORKDIR)
        2. Auto-generate plist — TCC policy compliant
        3. ProgramArguments — no /bin/zsh -lc wrapping ✓
        4. FD limit 65536, 3-layer defense ✓
        5. launchctl unload && load to re-register
Done. Next run: 05:00 daily

→ The full power of Skills shows up alongside CLAUDE.md evolution in Part 2. They’re a pair.

TIP 04 · SAFETYHarness engineering — hand over every permission, but hardcode the non-negotiables

Bottom line. I give Claude Code every permission, but the “these things, absolutely not” lines are pinned in code. I call this harness engineering.

If I’m going to hand Claude Code full permissions, I needed non-negotiable lines to do it safely. Blind safe mode wasn’t what I wanted. The essence was branching on whether I could actively review or not.

• Folder deletion rm -rf
• Git force push push --force
• Job deactivation launchctl unload
• Force-kill processes kill -9
• Docker container removal docker rm
• DB table drop DROP TABLE

It’s not locked all the time. It only blocks when the Telegram bot is running on its own while I’m away. When I’m sitting there reviewing, it lets things through.

#!/bin/bash
# .claude/hooks/pretool-destructive.sh

CMD="$1"
DANGER="rm -rf|push --force|launchctl unload|kill -9|docker rm|DROP TABLE"

if echo "$CMD" | grep -qE "$DANGER"; then
  if [ "$CLAUDE_AUTOMATED" = "1" ]; then
    echo "❌ Blocked in auto mode: $CMD"
    exit 1
  fi
  echo "⚠️ Dangerous command detected — allowed in review mode"
fi
exit 0

Every block fires a Telegram alert immediately. Thanks to the PreToolUse hook lifecycle, the command is stopped before it executes:

🚫 PreToolUse block
Command: rm -rf ~/Documents/Claude/Projects/automation/sessions/
Time: 2026-05-12 03:22:14 KST
Mode: auto ($CLAUDE_AUTOMATED=1)
Reason: DESTRUCTIVE GATE — folder-deletion pattern

To retry in review mode:
$ unset CLAUDE_AUTOMATED && {original command}

Situation                          $CLAUDE_AUTOMATED   Blocks            Allows
─────────────────────────────────────────────────────────────────────
Telegram bot auto-run               = 1                 6 risky commands   safe commands only
User review mode (not away)         = 0 or unset         (none)             risky commands included
Emergency recovery (explicit OK)    = 0 + user confirm   (none)             everything

The more permissions I delegate, the sharper the non-negotiable lines need to be. When I expand the automation surface, I scale the safety net at the same time — that’s the real order of operations.

→ The full auto/review split is coming in Part 4.

TIP 05 · IDENTITYThe 3-file pattern — building a Claude that writes like you

Bottom line. I registered 3 files — identity, voice, and don’t-write — as system prompts and got a real proxy of myself. Writing time dropped roughly in half.

I dropped 3 files into claude.ai Projects as system prompts. Every conversation now starts with identity, voice, and forbidden phrases baked in. After this, it really did feel like I’d built a proxy.

For voice.md I learned that samples teach better than explanations. So I just paste excerpts of my own writing.

# voice.md — Eddie's voice samples

## Sample 1 (failure retro)
The job died. I restarted it. It died again.
Same hook was hanging all three times.

## Sample 2 (tool review)
One Raycast shortcut is why I can't go back to Alfred.
Objectively they're a wash. It's just this one key.

anti-ai.md follows the same shape. Borrowing the “negative example” pattern from the Anthropic context engineering guide, I list only the phrases to block:

# anti-ai.md — phrases I will never write

## Marketing fluff
- "more efficiently"
- "optimal solution"
- "innovative ~"
- "surprisingly ~"

## Empty diagrams
- "[A] → [B] → [C]" arrow chains
- 4-quadrant matrices (when there's no real intent)
- Tidy analogies that explain nothing

## Absence of critical thinking
- "Overall, ~"
- "In conclusion, ~"
- "It seems likely that ~"

I baked the 3-file pattern into both claude.ai and Claude Code (via CLAUDE.md + Skills). Pinning it in both surfaces gave me results that matched or beat my own drafts. Writing one post used to take 30 minutes of hands-on time. Now it’s 12.

→ The full 3-file kit is coming in Part 5.

CODAThe 6th one is actually the most important

Two months in, I learned 5 things. Honestly it’s 6. The last one is meta.

#6 — I wasn’t perfect from day one either. I cycled request → result → review → revise a lot. At some point CLAUDE.md, Skills, slash commands, hooks, and the 3-file pattern had all grown up around me.

Disclaimer: personal 2-month operating log. No ads, no affiliates.