Anthropic Says This One Sentence Fixes Fable 5.2. I Tested It.

Anthropic’s new guide to prompting Fable 5.2 has a section called “Say who owns the code”, and one sentence from it is going around: “You are the owner and long-term maintainer of this codebase, and its design is up to you.” The guide says that on small coding tasks Fable merged duplicated code in under 10 percent of runs without that sentence and in about 75 percent with it. I added it to Claude Code and counted what changed.

📺 Watch the video: Anthropic Says This One Sentence Fixes Fable 5.2. I Tested It.

The sentence and where it goes

A system prompt is the standing instructions a model reads before your message. Claude Code already has one, so I added the sentence to it with --append-system-prompt, word for word from the guide. Everything else stayed the same in every lane: Fable 5.2 at medium effort, Claude Code 2.1.322 in -p mode on my Max 5x plan, one run at a time, with the lanes taking turns. The plan doesn’t bill per token, so the cents below are Claude Code’s own count at API list prices.

The planted copy

The repo is the shop till from my earlier videos. It checks ID for age-limited items: 18 for wine, beer, spirits, tobacco and lottery tickets, 16 for energy drinks. That rule lives in two places. The receipt exports it, and the self-checkout keeps its own copy, written a little differently.

The job was a delivery slip that marks which items the driver checks ID for. The task never mentioned the copies. The receipt’s rule is one import away, so the easy path is to reuse it and leave the second copy alone.

Hidden tests the model never saw check the slip byte for byte, check that the receipt and self-checkout print exactly what they did before, and count the copies of the rule that are left. One copy left means the run merged them.

Test 1: with and without the sentence

8 runs each Passed Merged the copy Typical output tokens Per passing run
Plain prompt 8 of 8 0 of 8 about 6,500 $0.60
With the sentence 8 of 8 8 of 8 about 8,000 $0.72

All eight plain runs saw the copy. Their own summaries say the self-checkout keeps its own copy, the two agree today, and they left it alone. With the sentence, every run made one small age file under ten lines and pointed the receipt, the self-checkout and the new slip at it.

The guide says output tokens rise 70 to 80 percent with the sentence. Here a typical run used roughly a quarter more.

The guide also warns about changes nobody asked for. No run changed how an old feature behaved, and the old receipts came out byte for byte the same every time. All 16 runs added a README section I never asked for, and three owner runs bumped the version number. Every summary listed both, so nothing was hidden.

Test 2: the rule that should keep it in check

The guide tried one more thing: the rule “No drive-by refactors, renames or cleanups.” earlier in the same prompt. Anthropic says Fable still merged in more than half the runs. I put that rule in front of the sentence and ran it 6 times.

All 6 passed and none merged. Every run left the self-checkout’s copy and gave the same reason: outside this task. It still wrote the README section every time.

On my till the rule cancelled the sentence completely, which is the opposite of what Anthropic saw. It’s one task, so I treat it as one data point.

Test 3: the rematch

In my last post Fable failed a real bug in Mock Service Worker 11 times, at medium, high and max. When the library resets, it throws an old generator away without closing it, so the cleanup in its finally block never runs. A careful owner locks up at closing time, so this time Fable got the ownership sentence plus the guide’s “Finishing work” block, which Anthropic says halves the unfinished sessions at medium.

It went 0 of 5, $4.76 in all. Every run fixed the restart the bug report described and missed the cleanup, the same two lines as last time, with its own tests green. Two runs did say which tests they hadn’t run, so the reports got more honest, but neither of them found the bug. That makes Fable 0 for 16 on this bug.

What I took from it

The sentence changed how Fable tidies and didn’t change what it notices. If you want to use it, here’s my checklist:

  1. Add the owner sentence when you want Fable to leave the code tidier than it found it.
  2. Leave it out when you want the smallest safe diff, like a hotfix, and don’t stack it with a no-cleanup rule hoping for a bit of both.
  3. Read the summary either way. Fable told me about every change it made.

This is one planted task with 8 runs a lane, and Anthropic’s numbers come from their own set of small coding tasks. I still don’t know why the rule won so completely here. If you run it on your own code, tell me what you got in the video’s comments.

Test the line yourself with which-claude

which-claude 0.6.0 adds prompt lanes, so you can run the same model with and without one line on your own task and see which one passes your check. example/owner is the till from this post with its hidden check (steps in the README):

which-claude "your task" --check "npm test" --models fable:medium --claude-code-prompt --prompt-lane owner="You are the owner and long-term maintainer of this codebase, and its design is up to you."

Code & video

If this was useful, subscribe to The Coderlog on YouTube for the next test.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top