Releases·Tool Release
Claude Style Patch: a drop-in Claude.md section to fix prose degradation in Opus 4.6+ and Fable
A heavy Claude user published a drop-in Claude.md patch on GitHub to fix prose degradation observed since Opus generations after 4.6. The patch came from iterative diagnosis of 'Claudish' writing tics, adding instructions to Claude.md and testing output changes. It matters to practitioners using Claude Code and Claude.ai for writing-heavy work.
273 engagement1 sourceTue, Sep 8, 2026, 03:30 AM
AI Overview
- Author reports prose degradation starting with Opus generations after 4.6 in both Claude.AI chats and Claude Code output
- Spec was iterated across Claude Code and Claude.ai over multiple Opus and Fable generations
- Cites Boris Cherny's public comments as evidence Anthropic agrees on Claude's communication shortcomings
Related
Products·launched
Behavioral harness to make LLMs behave like Opus 4.6
A set of behavioral instructions that can be applied to any LLM to reduce brashness and improve thoughtfulness and succinctness, inspired by Claude Opus 4.6. It is a markdown file that users can incorporate into their LLM's output style, solving the problem of annoying or overly verbose AI responses.
A set of behavioral instructions that can be applied to any LLM to reduce brashness and improve thoughtfulness and succinctness, inspired by Claude Opus 4.6. It is a markdown file that users can incorporate into their LLM's output style, solving the problem of annoying or overly verbose AI responses.
15 engagement1 sourceredditSun, Aug 9, 2026, 07:47 PM
Products·launched
Multi-LLM: review plans and code across 12 coding CLIs and many models
Multi-LLM is a plugin that lets developers review plans and code using multiple LLMs across 12 different coding command-line interfaces. It solves the problem of comparing outputs from different models and tools in one place, saving time for developers who want to leverage the best LLM for each task.
Multi-LLM is a plugin that lets developers review plans and code using multiple LLMs across 12 different coding command-line interfaces. It solves the problem of comparing outputs from different models and tools in one place, saving time for developers who want to leverage the best LLM for each task.
2 engagement1 sourcehackernewsWed, Jul 29, 2026, 07:26 PM
Products·launched
Claude Code writing standard based on The Elements of Style
A CC0 writing standard that adapts Strunk and White's The Elements of Style for Claude Code, making its responses and file prose concise, direct, and clear. It uses a root CLAUDE.md for defaults and a SKILL.md with the full rules for sustained dialogue and writing-heavy tasks.
A CC0 writing standard that adapts Strunk and White's The Elements of Style for Claude Code, making its responses and file prose concise, direct, and clear. It uses a root CLAUDE.md for defaults and a SKILL.md with the full rules for sustained dialogue and writing-heavy tasks.
100 engagement2 sourcesredditThu, Aug 20, 2026, 03:25 PM
Social·Community
Users report Claude Opus drifting from claude.md instructions, verbose comments
Developers report that Claude Opus increasingly ignores claude.md instructions mid-session, producing verbose, pedantic code comments with jargon and em-dashes. The behavior escalates with longer conversations and repository work, suggesting a reinforcement loop. One user measured replies showing high variance, with p90 length seven times the median.
Developers report that Claude Opus increasingly ignores claude.md instructions mid-session, producing verbose, pedantic code comments with jargon and em-dashes. The behavior escalates with longer conversations and repository work, suggesting a reinforcement loop. One user measured replies showing high variance, with p90 length seven times the median.
033398 engagement7 sourcesreddit, hackernewsTue, Aug 11, 2026, 08:05 AM
Social·Community
Phantom Cloudflare Worker deploys landing ~60s after every wrangler deploy. Anyone seen this?
A developer running a personal site on Cloudflare Workers reports that an unexplained deploy lands 40-80 seconds after every legitimate wrangler deploy. The phantom deploy is a real script_update/script_deploy sequence in the audit log, attributed to their own Gmail OAuth identity, and builds from the GitHub repo source rather than the local tree. It once reverted an unpushed local fix; the user suspects a forgotten agent session, stale CI runner, or something on Cloudflare's side.
A developer running a personal site on Cloudflare Workers reports that an unexplained deploy lands 40-80 seconds after every legitimate wrangler deploy. The phantom deploy is a real script_update/script_deploy sequence in the audit log, attributed to their own Gmail OAuth identity, and builds from the GitHub repo source rather than the local tree. It once reverted an unpushed local fix; the user suspects a forgotten agent session, stale CI runner, or something on Cloudflare's side.
1 engagement1 sourceredditWed, Sep 30, 2026, 07:09 PM