<aside> 🧰
The Six-Month Audit. Every instruction in your Claude setup was written to fix a problem an older model had. Nobody goes back and removes them once the model stops having the problem. This finds the ones that are now costing you. Free, nothing to sign up for — scroll and use it.
</aside>
<aside> 📖
How to read this. Section 2 is the whole idea in one picture. Sections 3 and 4 are the actual checks, and every single one quotes the line in Anthropic's docs that justifies it. Section 5 is three ways to run it depending on whether you use Claude Code. Section 6 is what happened when I ran it on my own setup, including the check that got it wrong.
</aside>
<aside> ✂️
When Opus 5 shipped, Anthropic's Claude Code team deleted more than 80% of Claude Code's own system prompt and the model performed better without it. The scaffolding had been written to compensate for weaknesses the new model no longer had. Their published advice is to try deleting your instruction file, your skills and your hooks every six months and see what the model does on its own.
</aside>
The uncomfortable part: the instructions doing the damage are usually the ones you were most proud of writing. "Always verify your work." "Double-check before answering." "Be conservative." Those were correct once. On the current models they are actively counterproductive, and Anthropic says so in writing.
The six-month audit loop
<aside> 🔄
Never hardcode the guidance. Fetch it live, every run. Model-specific advice changes every time a model ships. An audit with the rules baked in becomes exactly the stale artefact it was built to find.
</aside>
The three pages to read, every time:
| Page | What it gives you |
|---|---|
| Prompting best practices | The living cross-model reference. Start here |
| The page for your model | Linked from the page above. One exists per current model. This is where the delete instructions live |
| The prompt-engineering blog post | The general craft, and which older techniques are now optional |
Run this one first. It is where the wins are. Search your own files for each pattern.
<aside> 🔍
Quote the line you found, then quote the line in Anthropic's docs that condemns it. If you cannot find a supporting line, the verdict is KEEP. Never guess, never invent a source.
</aside>
| # | Search your files for | Why it goes now |
|---|---|---|
| 1 | Explicit verification steps. "Verify before done", "include a final verification step", "use a subagent to verify", "prove it works" | Anthropic: remove these, they cause over-verification and waste tokens with no quality gain. Explicitly includes legacy harness scaffolding that adds a separate verify step |
| 2 | Re-check instructions. "Double-check your answer", "re-verify before responding", "challenge your own work" | The model already catches and fixes its own mistakes. These compound with its own behaviour and add cost without improving results |
| 3 | Any rule telling the model not to think or not to reason | Increases internal XML tag leakage into visible output. Anthropic: remove it |
| 4 | Review prompts saying "only report high-severity issues" or "be conservative" | Followed literally, so you get told less. Ask for everything and filter in a separate pass |
| 5 | Vision or image workarounds you tuned for an older model | Re-validate. Anthropic: they may no longer be needed |
| 6 | Effort defaults carried over from a previous model | Re-run an effort sweep. low and medium now hold quality at a fraction of the cost and latency |
| 7 | XML tags and heavy role-play treated as mandatory structure | Once recommended, now optional. Keep where it genuinely separates mixed content, drop where it is ceremony |
<aside> 🏷️
Tag legend. Orange means delete it outright, the guidance is explicit. Blue means re-validate it against your own results rather than binning it blind.
</aside>
Newer models are more verbose and more eager than the ones your file was written for. That opens five gaps.
| # | What is missing | Why you need it now |
|---|---|---|
| 1 | An explicit conciseness instruction | Default responses run longer than prior models. Effort controls how much it thinks, not how much it says, so length has to be asked for directly |
| 2 | Written-deliverable length calibration | Files it writes to disk come out longer and padded with boilerplate unless you calibrate them |
| 3 | A narration cadence for agentic work | It announces what it is about to do. Describe the cadence you want. Positive examples beat prohibitions |
| 4 | A task-scope constraint | It expands scope and adds steps you did not ask for. Constrain it explicitly on narrow tasks |
| 5 | A subagent delegation cap | It delegates readily, which multiplies cost and time on small tasks. Cap it, or name which scenarios justify it |