Tuesday 10 November 2026
One prompt sentence got Fable 5.2 to merge duplicate code
Five bullets
The one thing
Anthropic's prompting guide for Claude Fable 5.2 says the model responds closely to who owns the code. In Anthropic's tests on small coding tasks, a prompt silent on ownership led the model to add to existing structure. It did that even when it meant repeating existing code.
The fix is one sentence from the guide. Add this line: "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." With it, the model merged repeated code in about half of tries. Without it, the rate was under a tenth.
Treat this as a lever, not a default. On code you own, more merging means less duplication to review later. On another team's code, adding to the existing structure may be what you want. The skill that grows is stating ownership and scope plainly in the prompt.
Decide whether you want cleanup or minimal change, then say so in the prompt.
Try this · 15 minutes
See whether one sentence changes how Fable 5.2 handles duplication
- Pick a small task in a file that already contains repeated code.
- Run it with Claude Fable 5.2 using your usual prompt. Save the diff.
- Revert. Rerun the same task with the guide's ownership sentence added to the prompt.
- Compare both diffs. Note whether the second run merged repeated code instead of adding more.
Stories
- TRYClaude Fable 5.2 becomes Claude Code's default Fable modelAnthropic released Claude Fable 5.2 and Mythos 5.2 , the same underlying model with different safeguards. Claude Code v2.1.321 makes Fable 5.2 its default Fable model, with 1M context at $8/$40 per million tokens. At High effort it roughly matched Opus 5.5 at Max on FrontierCode 2.0, for about half the cost per task. Anthropic still recommends Opus 5.5 as the daily driver, so Fable suits engineers with large, open-ended jobs.
- WATCHChatGPT Desktop lets remote agents use your SSH credentialsChatGPT Desktop forwards your local SSH agent to remote machines, according to a Hacker News report. An agent running on the remote host can then authenticate as you. The report says that means it can push to your private GitHub repos. It matters to anyone running Codex agents on remote machines from the desktop app.
- TRYLLM database code often borrows its crash safety from the librarySafe by Accident crash-tested 9,100 LLM-written solutions from ten models on 24 backend tasks. Of the database code that survived every crash under sqlite3's default mode, 54% was left half-updated in autocommit mode. Blocks written as "with conn:" looked like transactions but opened none. Asking for crash safety helped every model, so backend engineers should ask for it explicitly.
- TRYClaude Code patches an MCP OAuth token leak and a sandbox gapVersion 2.1.321 stopped redirects sending refresh tokens to another host. It also stopped a read-only Bash command running unsandboxed without a prompt.
- TRYdevvm gives each project its own VM for coding agentsNew VMs reach only allowed hosts by default. With devvm --clone, you fetch the agent's work and review it before merging.
- WATCHMeta's Muse Ember targets agent subtasks at Haiku-level pricesIt costs $0.10/$0.50 per million tokens on prompts up to 200k. It sits on Artificial Analysis's cost-per-task Pareto line.
- WATCHSourcegraph built a bot that picks which PRs need human reviewThe bot weighs the diff and the author's recent work in each component. Changes to authentication or sensitive data always need a human.
- WATCHSourcegraph ends Cody on January 31, 2027Sourcegraph points Cody users to Deep Search and its MCP integrations. It says these are not one-for-one replacements.
Safe to ignore
57 sources read. 107 items. 13 shortlisted. 5 bullets.