Open-source skill gives coding agents context-aware product copy, with an eval harness
Life_Assist5275 · reddit · 2026-10-28
- The author released an Apache-2.0 skill, adaptive-product-language, fixing a common agent failure: functional UIs shipped with generic, repetitive, or behavior-mismatched copy.
- Covers landing pages, onboarding, auth, forms, pricing/paywalls, CTAs, error/loading/empty states, and localization; it reads product domain, audience, market, real feature limits, pricing, and approved terminology, and routes compose/improve/audit/remediate/localize separately.
- Ships a lightweight eval harness measuring copy quality, unsupported claims, behavioral congruence, CTA expectations, template-like repetition, context size, and latency. Design principles: unsupported promises should fail independently of clarity; shorter isn't automatically better.
- Preliminary one-sample observations: Gemini 3.8 Flash output bytes 2,912→2,663 (-8.5%), Claude Sonnet 4.6 Thinking 3,289→3,034 (-7.8%); Codex showed no improvement. The author explicitly disclaims statistical significance.
- Asks the community how to test unsupported promises, measure cross-product repetition, and decide how much context a skill should load.
More from coding & agent
- Open-Source Skill Teaches AI Agents to Design and Produce Videos — songguoxiansen · 2026-10-28
- Not Everything Needs an Agent: Fixed Steps Are Recipes, Save Agents for Judgment — adamse · 2026-10-28
- Dev suggests self-hosting via Tailscale only: no open ports, tailnet as sole access path — vincesitzmann · 2026-10-28
- Letta's FrontierMemory leaderboard: Letta Dream tops at 73.3%, beating OpenAI and Claude agents — afurgs · 2026-10-28
- jackson-databind hit with CVSS 9.8 flaw CVE-2026-110879, bypassing prior fix — Kyrannio · 2026-10-28
- Researcher says Claude writes 50K lines of code for him per month — ethanCaballero · 2026-10-28