Turn SKILL.md Into a Trainable Artifact

... PLUS: Your AI Agent Is Missing Most of the Web

Author

Sumanth P
June 12, 2026

In today’s newsletter:

  • Your AI Agent Is Missing Most of the Web

  • SkillOpt: Turn SKILL.md Into a Trainable Artifact

Reading time: 5 minutes.

Most web search APIs are built for humans skimming ranked results. When an AI agent uses them, it gets the top few pages and stops. Everything else goes unseen.

For a human browsing a topic, that's fine. For an agent making a risk or market decision, missing 80% of relevant events is not a minor gap.

NewsCatcher builds two APIs to fix this: CatchAll for web search and a News API for news data. Both are built around recall, finding everything relevant, not just the most prominent results.

Finding What Other Tools Miss

Standard deep research tools are limited to the first few pages of results. CatchAll scans 10,000+ pages per minute across a 2B+ page index, enriches every result with NLP tags, and returns structured machine-readable output instead of raw links.

Where Recall Actually Matters

The use cases where missing events has real consequences:

  • Risk and threat intelligence: a missed lawsuit filing or regulatory action is not a small gap

  • Supply chain monitoring: factory strikes and port disruptions appear in regional sources most APIs never index

  • Regulatory tracking: policy changes surface in government and municipal sources before they hit mainstream news

When You Need News API Specifically

Where CatchAll covers the open web, the News API is specifically for news. It pulls from 140,000+ sources across 100+ countries, articles arrive within minutes of publication, and the archive goes back to 2019.

What separates it from a raw news feed is the enrichment layer. Entity resolution maps every article to the exact company or person you're tracking, not just a name match. Events intelligence tags key signals across 100M+ data points daily. Localized filtering goes down to city and region level.

Real-World Results

  • UC Berkeley found 20% more articles captured versus alternatives in benchmark testing

  • HCOB uses it for credit risk assessment, rating it best on availability, quality, and regional focus

  • A global electronics manufacturer attributed seven-figure annual cost savings to supply chain monitoring built on the API

Fine-tuning an LLM is often expensive and slow. SkillOpt offers a different solution: instead of training the model, it trains the document the model reads.

SkillOpt is a text-space optimizer for agent skills that treats your SKILL.md file as the trainable parameter and improves it through a feedback loop. No GPU. No weight updates. The model never changes.

How It Works

A separate optimizer model watches the agent run on training tasks. It reads the rollout trajectories, identifies where the agent failed, and proposes bounded edits to the SKILL.md: add a section, delete a line, replace a passage.

Each proposed edit gets tested against a held-out validation split. If it strictly improves performance, it's accepted and becomes the new baseline. If it doesn't, it's rejected and stored as negative feedback that shapes future proposals.

The deep learning analogy the paper draws is intentional:

  • Rollout batch is your training data

  • Edit budget is your learning rate

  • Validation gate is your validation set

  • Rejected-edit buffer is your negative feedback signal

The optimizer runs offline. The deployed artifact is just a static SKILL.md file.

The Results

Tested on GPT-5.5 across six benchmarks and three harnesses:

  • Direct chat: +23.5 points average over no-skill baseline

  • Codex: +24.8 points

  • Claude Code: +19.1 points

Best or tied-best on all 52 evaluated cells against Trace2Skill, TextGrad, GEPA, EvoSkill, hand-written skills, and one-shot LLM-generated skills.

What's striking is how little it takes to get there. Most gains come from 1 to 4 accepted edits. The final skill stays compact at 300 to 2000 tokens. SpreadsheetBench jumped from 41.8 to 80.7. OfficeQA from 33.1 to 72.1. One single accepted edit gave OfficeQA a 39-point gain.

Skills Transfer

Optimized skills aren't locked to the model or harness they were trained on.

A SpreadsheetBench skill trained inside Codex transferred to Claude Code with a +59.7 point gain. Skills trained on GPT-5.4 improved every smaller GPT variant tested. ALFWorld went from 70.9% to 85.8% with a skill trained on GPT-5.4-mini.

The SKILL.md file is portable. Train once, deploy anywhere.

That’s all for today. Thank you for reading today’s edition. See you in the next issue with more AI Engineering insights.

PS: We curate this AI Engineering content for free, and your support means everything. If you find value in what you read, consider sharing it with a friend or two.

Your feedback is valuable: If there’s a topic you’re stuck on or curious about, reply to this email. We’re building this for you, and your feedback helps shape what we send.

WORK WITH US

Looking to promote your company, product, or service to 200K+ AI developers? Get in touch today by replying to this email.