Building an Article Research Stack with AI Agent Skills

An agent that writes articles no longer just summarizes search results. With the Agent Skills format, research expertise can be packaged into files an agent loads only when needed. I recently rebuilt this blog's article research pipeline using that approach. Here is the full account: what worked, what I rejected, and the security numbers nobody talks about.

Key Takeaways

• Agent Skills is now an open standard: one SKILL.md file works across agent platforms.^[1]
• For article research, three core capabilities are enough: a structured web research workflow, a citation ledger, and a fallback ladder for blocked pages.
• Third-party skills look tempting, but a February 2026 security audit found that 36.82% of 3,984 public skills carry at least one security flaw.^[3]
• All-in-one article-writing plugins are usually too heavy; adopting their editorial patterns is safer.
• The biggest cost is not installation. It is LLM tokens per research job, controlled by declaring the research scope up front.

What Agent Skills Are

Agent Skills are folders of instructions, scripts, and resources that an agent can load dynamically when a relevant task appears. Anthropic introduced them on October 16, 2025, then published them as an open standard on December 18, 2025 so the same skill can run across many platforms.^[1]

Technically, a minimal skill is just a directory containing a SKILL.md file with YAML frontmatter (a name and a description) plus markdown instructions.^[2] The name is capped at 64 characters and the description at 1024, and the description is what tells the agent when to activate the skill.^[2]

Why the Format Fits Article Research

The key mechanism is progressive disclosure. At startup, the agent pre-loads only each skill's name and description as the first level. The full SKILL.md body is read only when the task matches it, as the second level, and bundled reference files load on demand as the third.^[1]

The practical effect: dozens of research skills can sit installed without weighing down the context window. Irrelevant skills rest in the index at a cost of tens of tokens each. That is why a tiered research stack can run with no extra daemon and no new API keys.

The Minimal Stack That Actually Works

From building a real article pipeline, three core capabilities deliver most of the value:

1. A structured web research workflow: split the topic into research lanes (definitions, current state, official docs, implementation evidence, limitations, counterevidence), then stop once the load-bearing claims are supported.
2. A citation ledger: every URL is registered at retrieval time rather than retyped from memory; verbatim quotes attach to their sources; the draft is mechanically verified against the ledger before release. This very article was written that way.
3. Blocked-page recovery: Wayback Machine, archive.today, server-side rendering, an API pivot, then a real browser, with provenance discipline. Archived snapshots are cited as snapshots, never as the live page.

Optional add-ons: academic paper search, RSS monitoring for recurring ideas, and a writing-style audit. The style audit runs only after the facts are locked, because polishing sentences must never shift technical meaning.

Lessons from Evaluating Third-Party Skills

I evaluated four popular candidates before deciding. Not all of them earned a place:

• An all-in-one blog suite (30+ sub-skills, an orchestrator, a five-gate delivery contract) is the most complete in its category, but it is built for one specific runtime with its own plugin, agent, and script structure. Installing it wholesale risks duplicating a publishing pipeline that already has its own gates. My decision: adopt the patterns, skip the package.
• A deep-research skill with an explicit ledger is format-compatible since it only requires Python per its metadata.^[5] It shines for contested or version-sensitive topics but adds artifact overhead. Worth it only if heavy research articles are a routine.
• An academic research pack targets 60 verified references per research run and depends on a paid third-party CLI.^[4] For ordinary blog posts that is clearly wasteful.
• Modular SEO skills (content brief, keyword clustering, meta optimizer) have solid instructions, but parts of their flow assume SERP and keyword-volume data that is not automatically available. Without that data, the output degrades into estimation.

Patterns worth stealing from editorial skills: require a brief before writing, open with a hook that promises specific value, and favor concrete substance, for example a number instead of the word "many".^[7] Placing keywords in the title, meta description, and opening paragraph without keyword stuffing is a baseline checklist shared across SEO skills.^[8] For quotable direct-answer blocks, a 50 to 70 word limit in a factual third-person tone proved easy to run.^[6]

The Dark Side: Third-Party Skill Security

This is the part most people skip. The first comprehensive security audit of the Agent Skills ecosystem (February 2026) scanned 3,984 skills from two public directories. The findings: 13.4%, or 534 skills, contained at least one critical-severity issue, from malware distribution and prompt injection to exposed API secrets. At any severity level, 36.82%, or 1,467 skills, carried at least one flaw.^[3]

The growth rate is extreme: daily submissions jumped from under 50 in mid-January 2026 to over 500 by early February, a 10x increase within weeks.^[3] Growth that fast makes manual curation nearly impossible. The format's own authors recommend installing skills only from trusted sources and auditing their contents, including code dependencies and any instruction that tells the agent to reach out to external network services.^[1]

The Rules I Run By

1. Internal skills first. The agent's built-in abilities (web search, page extraction, terminal, files) cover most article research needs. Extra skills are only for genuinely recurring procedures.
2. Read before installing. The SKILL.md, the scripts folder, and the references get read first. Skills that ship executable code get extra scrutiny.
3. No daemon if possible. Markdown-plus-local-script skills are safer and cheaper than always-on services.
4. Ledger first, writing second. Citations are recorded at retrieval; drafts are mechanically verified before entering the publishing pipeline.
5. Style last. The "sounds too much like AI" audit runs only after facts are locked, in detect or manual-edit mode, never as an automatic rewrite.
6. Declare the research scope. The default is bounded: enough sources to support the claims, not an academic 60-reference sweep.

Frequently Asked Questions

Skills and SEO tools, which replaces which

Neither replaces the other. Skills organize the agent's procedures; SERP data, keyword volumes, and rank tracking still require the data tools. What changes is the speed of executing research procedures and the consistency of citations.

The real cost of this stack

The skills are free and daemon-free. The variable costs are LLM tokens per research job and, if used, external data APIs. Keep them under control by declaring an explicit research scope per article.

Is it safe to install skills from the internet

The risk is real; see the audit numbers above.^[3] Safe practice: trusted sources, read the contents before installing, avoid skills that request excessive access, and pin versions.

The risk of being detected as AI writing

If the article only summarizes search results, yes. If the research is evidence-first, with primary sources opened, claims mapped, and counterevidence checked, plus a genuine authorial angle, the piece has earned its place and usually competes well.

The Takeaway

The Agent Skills format changes how content workflows get assembled: expertise becomes a file that can be audited, shared, and loaded on demand. For article research, the win is not the number of skills installed but the discipline of the process. Sources are opened before they are cited, claims are mapped to evidence, and third-party skill security is taken seriously from day one.

Sources

• Anthropic Engineering: Equipping agents for the real world with Agent Skills
• Agent Skills Specification (agentskills.io)
• Snyk ToxicSkills: security audit of the Agent Skills ecosystem (February 2026)
• K-Dense research-lookup SKILL.md
• deep-research SKILL.md (B143KC47)
• ccforseo content-brief SKILL.md
• content-and-copy SKILL.md (rampstack)
• seo-blog-writer SKILL.md (skillsdirectory)