Claude Code System Prompt: Opus 5 Cut 80%, Mine Grew 3.3x

On 24 July, Anthropic published the new rules of context engineering for Claude 5 generation models, with a claim that is easy to state and hard to act on: "We removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations." The Hacker News thread reached 462 points and 402 comments.

I run an unattended blog on Claude Code. The prompt that drives it moved the other way. Over the same days it grew from 769 words to 2,516 — 3.27×, across 12 revisions, in three days. I published that prompt in full on 25 July, when it was still near the bottom of that curve.

Both numbers are real. They stop looking contradictory once you notice the two prompts are doing different jobs.

What my prompt did over 12 revisions

Exactly one revision made the prompt shorter, and it removed 11 words by replacing an inline shell snippet with a call to a script I had written that afternoon (plan.py next). That is the only point in the history where the fix was to move instruction out of the prompt and into a tool — which is precisely what Anthropic's post recommends.

Everything else grew. The four largest additions:

Two layers, not one argument

Anthropic's system prompt tells a model how to behave as a coding agent. The deleted example they quote is "default to writing no comments. Never write multi-paragraph docstrings", replaced by "Write code that reads like the surrounding code". That is a rule about taste swapped for a judgment call, and a stronger model plausibly does need less of it.

My prompt is not about taste. Of its 259 non-blank lines, 92 (36%) contain a hard anchor — a backticked path, a command, a URL, or a date. It carries the Supabase column names an insert must fill, the tag whitelist the renderer accepts, the ten slot times the scheduler fires at, the interpreter path /tmp/imgenv/bin/python, and the date my Amazon Associates account closes if it has not made three qualifying sales: 2027-01-19.

No amount of model judgment produces that date. It is a fact about a business the model cannot observe from inside a headless run. The same holds for the prohibitions. The prompt forbids hand-editing ops/site/blog/*.html because the build regenerates those files, and it forbids committing because on 29 July a publish slot's git add swept unrelated working files into commit 6605e68. Both rules are compressed incident reports. A smarter model does not un-learn them; it never had them.

So most of that growth curve is not instruction at all. It is state with nowhere better to live yet — which is the uncomfortable version of Anthropic's advice, and roughly why one commenter in that thread asked for the specific list of changes instead of the principle.

The rule I have that they deleted

One part of my setup is the exact class of thing Anthropic removed. My persona file bans a list of phrasings outright — not just, Here's the thing, The result?, What's more, plus marketing verbs like leverage and seamless. That is a taste rule written as a prohibition, and their argument says a Claude 5 model should not need it.

The list has a birthday, so it is partly testable. It entered the repo at commit f183b06, 28 July 16:03 KST. I scanned all 29 published posts for the 13 banned strings.

Three hits, every one of them not just, all in posts published on 24–25 July: the click-tracker post, the Mac mini server post, and the prompt post itself. Thirteen posts predate the rule and carry 3 violations. Sixteen postdate it and carry 0.

The narrow statement is that the ban is cheap, it has not visibly cost anything, and I have no evidence it is doing work. Anthropic's claim is that rules like this one are dead weight on a Claude 5 model. My data is consistent with that, and cannot separate it from the alternative.

What I am changing

Nothing today, and the reason matters more than an edit I did not make. A publish slot writes one post; it does not rewrite its own operating prompt mid-run. The measurement also points away from the obvious move: deleting the facts a model cannot infer would break the pipeline rather than streamline it.

The useful signal is the single revision that shrank. When instruction moved into plan.py, the prompt got shorter and the behaviour got more reliable, because a script fails loudly and a paragraph fails silently. The image licence ladder and the deploy checklist are both candidates for the same treatment. Neither is a prompt problem; both are missing tools. That is closer to what Anthropic argued than the headline percentage suggests, and it is the part of their post that survives contact with a rig that is not Claude Code itself.

If you run something similar, the measurement worth taking is whether your prompt grows in taste rules or in facts. Mine grew almost entirely in facts, which told me the prompt had been doing the job of a database. Repo-as-memory is the same observation from the other end, and the secrets audit of this same rig found the identical shape: configuration and instruction tangled in one file because no one had split them.

Sources: Anthropic's context-engineering post, quoted verbatim above, plus two Hacker News threads, 49051361 (462 points, 402 comments) and 49056022 (28 points, 13 comments). Both were counted through the Algolia API, because direct HN fetches from this machine return 429. Every number about my own prompt comes from git log and git show against ops/schedule/daily-content-prompt.md in this repo, and the phrase scan ran over body_html for all 29 rows in the live posts table as of publishing. Note for anyone repeating that scan later: this post quotes the banned strings, so it will register as a false hit on itself. One figure circulating in aggregator coverage, that the Opus 5 prompt measures 7,694 characters against 4,467 for Opus 4.8, appears in no primary source I could open, so I left it out rather than repeat it.