XNews

Agent tooling, skills, prompt transfer and evaluation

Agent tooling, skills, prompt transfer and evaluation

Developer communities published multiple agent tools and frameworks for retrieval, skills and cross-model prompt portability. - OpenNote's Index: low-latency, multi-tenant retrieval built on Postgres + pgvector with global Cloudflare edge, P99 <300ms, hybrid search and careful caching - Index ingestion pipeline uses chunk hashing and partial re-embedding to save compute - Claude 'skills' are prompt-driven meta-tools: SKILL.md with YAML, allowed-tools, model fields and Markdown instructions; discovery via a Skill meta-tool - Skills change runtime context without remote code execution and enforce minimum permissions - PromptBridge framework translates tuned prompts across models without retraining using MAP-RPE calibration and mapping pairs - Agent evaluation proposals include expert scoring, model scoring, direct comparisons and expert-user panels; ordinary-user scoring discouraged for core metrics - PromptBridge and skills aim to reduce model lock-in and speed cross-model migration

Key insights

OpenNote Index optimizes agent retrieval with Postgres and edge runtime

OpenNote published Index, a document retrieval system designed for agents and personal knowledge bases. Index uses Postgres (pgvector + pg_trgm + tsvector) rather than specialized vector DBs to exploit fast metadata filtering, supports hybrid keyword + vector search, and runs on Cloudflare Workers for global low latency. The ingestion pipeline hashes chunks to avoid unnecessary re-embeddings, and a three-layer cache plus precise invalidation maintains freshness while keeping P99 latency under target thresholds.

Claude 'skills' are prompt-driven modular behaviors for safe complex workflows

An analysis of Claude's skill system shows skills are manifest as SKILL.md directories containing YAML metadata and Markdown instructions. Skills are injected as high-priority contextual prompts, pre-authorizing minimal tools and changing model behavior without granting code execution. The system supports discovery via a Skill meta-tool, runtime injection, and hot updates. The architecture prioritizes safety through least-privilege allowed-tools and pure prompt-driven behavior.

PromptBridge automates prompt translation across models without retraining

PromptBridge, described in an Accenture paper, provides a zero-training framework to map high-quality prompt formulations from a source model to a target model. It uses small calibration sets (MAP-RPE) to generate paired optimal prompts and then learns a mapping to translate future prompts, enabling prompt portability across models with minimal human rework.

Agent evaluation should prioritize expert scoring and comparative model runs

A proposed agent evaluation pipeline recommends deriving an expert-based scoring rubric, collecting expert-graded case sets, running multiple models for model-based scoring comparisons, and validating results with expert users. The recommendation discourages relying on non-expert public scoring for technical agent metrics, while still acknowledging NPS value from general users.

Source posts