Anthropic documentation reveals that simply telling Claude Fable 5.2 the codebase is its own raised its rate of merging duplicated code from under 10% to roughly 50% in small coding-task tests. The finding shows that ownership framing strongly steers whether the model conservatively extends existing structure or actively refactors it.
Prompt Engineering
提示词工程
Practice of designing prompt content and wording to steer LLM behavior and output quality.
Curated Intelligence
Google Research has open-sourced Agile Deliberation, a framework that lets users define fuzzy 'gray area' concepts before an AI classifier learns them. The system splits a concept into in-scope and out-of-scope subconcepts, surfaces borderline samples for user labeling, and has Gemini rewrite the definition as a VLM prompt. Across 18 live sessions it beat zero-shot Gemini 2.5 Flash by 10.5 F1 points on average, with code released under Apache 2.0.
The article argues that prompting is fundamentally programming, explaining why software engineers see massive productivity gains from AI while non-technical roles like marketing and management struggle to realize promised benefits. It suggests natural language programming currently accelerates those who already think like programmers rather than democratizing coding for everyone.
Own HackerTube has launched an AI-powered video curation tool inspired by Andrej Karpathy's concept. It allows users to define preference standards in natural language, automatically filtering high-quality content from followed YouTube channels and learning from feedback to generate personalized video sites.
DeepSeek is promoting 'Creator Mode' in its open-source agent app Harness, enabling users to generate working plugins from plain-language requests where the AI writes, installs, and tests them. Although the feature has existed since August, the release of a demo video and prompts aims to leverage Harness's 268K GitHub stars to grow the ecosystem, pending validation of plugin reliability beyond curated demos.
Industry commentary suggests that Manus's Electron client and custom LLM API expose its prompts, skills, and harness logic to reverse engineering, significantly lowering technical barriers. While cloning is feasible, brand trust and deep product decision-making remain key differentiators, making simple replication unlikely to succeed.
Cheng Jiawei from Singularity Chuanyue suggests that recording human performances for micro-expression and motion reference is more stable than prompt engineering alone in AI video generation. His team uses a workflow of fixed scripts/prompt templates with Agents for batch generation, though narrative depth, spatial continuity, and asset licensing remain key bottlenecks.
Addy Osmani from the Anthropic Claude Code team advises users to stop 'over-prompting' as newer models like Opus 5.5 improve. He recommends using single clear instructions combined with automated acceptance criteria and appropriate reasoning effort settings to allow the model to execute tasks more independently.
Startup Dalet Labs raised a $7M pre-seed led by Lightspeed to build an AI-generated shopping feed. The system uses LLMs to extract attributes from liked images for diffusion prompts, reducing hallucinations and enabling self-growing content without photoshoots.
Rust off Rails introduces a 'zero-code framework' concept, using 60 Markdown files to define stack conventions and guide coding agents in generating compliant full-stack Rust web apps. It aims to solve architectural inconsistency in AI programming by letting developers focus on requirements rather than boilerplate.