✂️ Ablation Audit
Anthropic deleted more than 80% of Claude Code's system prompt when Opus 5 came out. This course teaches you to do the same with your CLAUDE.md, your skills and yours hooks — diagnosing line by line, cutting safely, and proving it through testing that the minimum version didn’t make things worse.
“Every line in your configuration is guilty of complexity until proven useful.”
But the goal isn't to cut as much as possible. It's to maximize quality + autonomy + verifiability ÷ complexity.
Your course progress
Your progress, questions, and notes are saved in your own browser. Nothing is uploaded to any server.
Why the config that worked became dead weight
Every instruction you wrote is the fixing a specific model’s weakness. When the next model no longer has that weakness, the instruction doesn’t disappear on its own: it keeps being reread every time it runs, taking up context and limiting the model. Ablation is the method for finding out which lines still justify their cost—you remove them and measure instead of guessing.
What to look at: in the left stack, the gray blocks are fixes for models that are no longer in use—and yet they’re reread on every run. The amber blocks are the ones the model can’t infer it on its own: these survive the ablation intact. On the right, notice the green block: verification is the only item that normally enters into an audit instead of going out—that’s the costliest gap of all.
What you’ll know how to do at the end
✓ You’ll leave here able to
- ✓Classify any instruction in your config into a category and a decision
- ✓Rewrite “do A, then B” as an objective + guardrails + criteria + verification
- ✓Run the skill
audit-ablacaoand read the 10-section report - ✓Move the procedure from
CLAUDE.mdfor skills loaded on demand - ✓Prove with A/B/C testing on real tasks that the minimal version didn't get worse
- ✓Keep the config auditable — knowing why each remaining line survived
✗ What this course does NOT do
- ✗Doesn't tell you to delete everything on impulse—diagnose first, then test
- ✗Don’t treat “smaller” as synonymous with “better”: short prompts without context break things
- ✗Doesn't teach Claude Code from scratch—it assumes you already use it and have a config
- ✗Don't cut security, compliance, integrations, or sources of truth
- ✗Don't bring the instruction back as a precaution—only after a real, repeated failure
The 4 tracks
🧬 Why delete
The 80% cut at Anthropic, why configuration gets stale, and the four-step ablation method.
🔍 How to diagnose
The 10 categories, the 6 decisions, the 7 questions—and turning a cake recipe into a criterion with verification.
🩺 Audit for real
Install and run the skill audit-ablacao, read the 10-section report and turn the Top 10 into safe cuts.
📊 Prove and maintain
The A/B/C test on real tasks, the rule for reintroducing instructions, and the cycle that keeps the buildup from returning.
Do you recognize any of these?
Anyone who’s used Claude Code for a while accumulates the same layer of sediment. If two of these symptoms sound familiar, the course pays for itself in time saved by Track 3.
| Symptom | What it usually is |
|---|---|
CLAUDE.md of 300 lines with dated exceptions | corrections for models that are no longer in use |
the same rule in the CLAUDE.md and inside 3 skills | redundancy that creates conflict when one of the copies changes |
| skill that teaches “how to think” in 400 lines | generic reasoning the model already does on its own |
| rigid 12-step process | micromanagement that blocks the better path |
| two rules that contradict each other | the model picks one, and you don't know which |
| nothing said how to check that turned out right | the costliest gap of all |
Who it's for (and what they need first)
Prerequisites
Claude Code installed and in use; at least one CLAUDE.md (global or project-level) and/or custom skills; basic familiarity with git and the terminal. This isn’t an introductory Claude Code course.
For whom
Anyone who uses Claude Code day to day and has accumulated config: from advanced users to developers. If you’ve never written a CLAUDE.md, start with a course on how to use it and come back here later.
💡 How to study
The common thread is the your config: it’s audited in Track 3, trimmed, tested, and put into a cycle in Track 4. Mark each section as read, highlight passages, and mark “I have a question”—everything returns in the panel My journey.
What you’ll deliver at the end
Four deliverables, all about your own configuration:
CLAUDE.md before/after — with the diff.Approval: all 4 items are present; every applied removal has a cited passage, a risk, and a way to test it; every restored instruction has a repeated failure recorded; and the final version has at least one objective check where there wasn't one before.