SShortSingh.
Back to feed

Tests find 11 of top 20 agent skills clearly helped; the most-installed one hurt output

0
·1 views

A DriftProof test ran the 20 most-installed agent skills from skills.sh with and without each skill loaded, using two Claude models and a Claude Haiku judge. Eleven skills measurably improved answers on both models, while the most-installed skill, find-skills, lowered scores on both. Ten of the 40 comparisons were inconclusive because answers varied too much between runs to draw a clear conclusion. The evaluation used three self-written text-only tasks per skill, excluding tool calls, media, and deployed services. The authors note that installs measure popularity rather than quality and have published their receipts and task suites for public review.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Hosted Checkout Breaks Purchase Analytics; Webhooks Restore Attribution

A developer building subscriptions for Belle AI found that a standard GA4 purchase event on the thank-you page recorded almost no sales. Hosted checkout sends users to the payment provider's domain, so they often never return to the site, and revenue appeared in the payment dashboard but not in analytics. To fix this, the developer stored first-touch UTM parameters and referrer data in localStorage, which are preserved across later visits. At checkout, that attribution data and GA4 identifiers were passed into the order metadata, which the payment provider returns in its webhook. The webhook handler then sends a purchase event to GA4 via the Measurement Protocol, reusing the stored client and session IDs so sales are credited to the correct traffic source.

0
ProgrammingDEV Community ·

LLM API list prices steady in mid-October as Chinese models trade at about a fifth of US prices

The TANDIU weekly data note for October 13 to 19, 2026 tracks official list prices for 69 AI models and reports that no vendor changed an official list price that week. Chinese models are priced in RMB, so their dollar equivalents shifted only with the exchange rate. The blended average is about ¥56.48 per million tokens for US models versus ¥10.79 for Chinese models, a gap of roughly 5.2 times. DeepSeek V4.1 Flash drew attention on social media, with its input and output tokens costing about 1/13 and 1/17 of Claude Opus 5.5's at list price. The same period's funding tracker logged over 150 rounds, including Ramp's $1.85B Series G at a $60B valuation.

0
ProgrammingDEV Community ·

Why a two-node Proxmox cluster freezes, and how a QDevice vote fixes it

A two-node Proxmox cluster can lock itself up because losing either node leaves the survivor with exactly half the votes, which is not a majority. Without quorum, the cluster filesystem at /etc/pve goes read-only, blocking guest starts, config changes and backups even though the surviving node is healthy. Adding a QDevice, a small external voter that can run on a Raspberry Pi or similar always-on Linux box, provides a third vote so the survivor keeps quorum. The QDevice must run on separate hardware, since placing it on a cluster node would remove two votes if that node fails. Unstable links, such as Wi-Fi with high jitter, can also make a cluster appear broken by causing nodes to repeatedly drop out and rejoin.

0
ProgrammingDEV Community ·

Choosing Between Vercel and Cloudflare for Low-Cost Website Hosting

A solo developer working on a micro-SaaS product shares lessons on keeping hosting costs minimal while avoiding unexpected cloud bills. The piece compares traditional VPS setups, which demand manual server management, with serverless platforms that handle operations and scale automatically. It argues that early-stage founders should default to serverless to preserve their time for building and validating products. The author also notes that workloads needing persistent processes, heavy computation, or custom binaries are better suited to VPS or container hosts. The article then turns to a comparison of Vercel and Cloudflare Workers, as the author's preferred platform choice.

Tests find 11 of top 20 agent skills clearly helped; the most-installed one hurt output · ShortSingh