01 · COMPACT/compact — press it on the alert, judge by the next answer

Bottom line. I learned /compact fires when the context alert appears. The way I see it, it summarizes and compresses the conversation to free up token space. It is not lossless.

My pattern — see the alert, run /compact, ask one question, and if it looks wrong I /clear and restart. No blind trust.

I think being honest about limits is the better move. I accepted that compaction is not lossless. Right after a critical code change, I now habitually restate the key facts before compacting.

# See the alert → /compact (first attempt)
[Claude] context window 80% used. /compact recommended.
$ /compact
[Claude] Compaction done — 60% token space recovered.

# Immediate one-line check after compacting
$ "Show me the contents of [key file] you just worked on"
# If clean, continue. If loss detected, /clear and restart

02 · AGENTImportant work goes to a separate agent, not main

Bottom line. The second pattern, in my view, is role separation. I run main Claude as the coordinator and throw detail work to sub-agents.

I see two payoffs. Main context stays protected + only the final output of the detail work surfaces back to main. The biggest win is that reading a huge document once doesn’t eat main context.

I keep 4 custom agents as one-pager markdown files under .claude/agents/:

On top of that, there’s the /review slash command and the weekly weekly-blog-review job (6-AI panel for post quality). This was the single biggest token-saving move.

# Don't ask main Claude
$ Task(opportunity-ranker, "rank the next 10 post idea candidates")
# → sub-agent works in a separate context, only the result returns to main

# Main plays coordinator only
# Heavy analysis / 6-AI panels / benchmarking all go to separate agents

03 · CLEAR/clear — not compression, full reset

Bottom line. Third, I actively use /clear.

A typical example for me — finishing real-estate blog work and switching to debugging the trading system. I realized mixing the two contexts causes incidents.

I see the two tools as different roles. /compact is compress, /clear is full empty. After clear, the first message reloads only CLAUDE.md and I rebuild context from there.

This was a bit counterintuitive: clear isn’t wasteful — NOT clearing is more expensive in most cases. Without clearing, every turn has to spend tokens replaying the entire prior context.

# /clear if you hit ANY of these
[ ] Complete domain switch (real-estate → trading system, etc.)
[ ] Answer feels off even after /compact (compression loss)
[ ] Conversation getting long, answers slowing down (context bloat)
[ ] Moving to another blog · project (separate CLAUDE.md folder)

# 1+ above → don't hesitate, /clear
# Clear, reload CLAUDE.md, restart clean

04 · SPLITSlim CLAUDE.md, move the rest to side files

Bottom line. The fourth is the CLAUDE.md diet.

The reason is clear to me. CLAUDE.md is auto-loaded every session. Every line in there spends tokens every time. The conclusion: anything you rarely look at has to come out.

My actual split structure:

I split old incident logs into CLAUDE-archive-2026-04.md and kept only a one-line index in main. The main file had been swelling past 1,000 lines — you can feel the relief. To me, this is the heart of the main-slim pattern.

# Main CLAUDE.md (currently ~80 lines)

## 5. Incident log (learnings)

| Date | Incident | Action |
|---|---|---|
| 2026-05-09 | 4 publish failures + telegram loop | 4 self-recovery layers added |

> 10 incidents from 2026-04-14 ~ 2026-04-18 moved to `CLAUDE-archive-2026-04.md` §A
> 8 incidents from 2026-04-21 ~ 2026-05-04 moved to `CLAUDE-archive-2026-05.md` §A
> Main keeps only post-05-05 incidents

05 · SKILLS3 skills — loaded only on trigger

Bottom line. The last one, in my view, is skills. I keep 3 custom skills under automation/.claude/skills/.

I couldn’t measure the split effect quantitatively. My read: inline = main CLAUDE.md swells past 1,000 lines and gets fully loaded every session. Split = main keeps a one-line index, and SKILL.md only loads when triggered. I estimate ~30% main-token savings.

---
name: blog-style-guide
description: 9 writing patterns + reviewer gate thresholds for blog post
  writing · review · publishing in the automation project. Trigger on
  ediblog · jessinvestment · luna-pick · jesslab publishing
  tasks or post quality review requests.
---

# Automation project blog style guide
...

# When Claude matches keywords like "write this post" / "review" / "before publishing"
# Auto-trigger → load SKILL.md → apply 9 patterns

CODAIf I had to nail it in one line

All 5 only stuck after I broke them at least once each. I’d recommend looking at what info you actually need every session before memorizing tools. That, I realized, is the essence of token saving.

Disclosure: Personal 3-month retro. No ads, no affiliates.