Anthropic Agent Skills documentation describes skills as reusable, testable instruction files, not one-shot chat prompts. In practice, editors report the same comparison failures after every agent run: mismatched tier names, pricing claims without URLs, and “best for” lines that drift from the research packet. A comparison page skill encodes that judgment so agents and humans share one standard before any draft stage.
TL;DR
- A comparison page skill holds inputs, falsifiable rules, examples, and eval hooks, not a template HTML file.
- Parameterize competitor pair, ICP, risk tier, and citation strictness.
- Golden tests must fail on wrong tier names and uncited claims before production.
- Wire the skill into the generate comparison pages ai agents pipeline at draft and QA stages.
- Version and promote the skill with version marketing workflows discipline.
What a comparison page skill is (and is not)
The skill is a versioned artifact agents invoke when rendering side-by-side copy from structured research. It is not a mail-merge template. It is not a 2,000-word prompt pasted into a canvas.
| Artifact | Holds judgment | Survives handoff | Has golden tests |
|---|---|---|---|
| Template HTML | Layout only | Yes | No |
| Chat prompt | Session tone | No | Rarely |
| comparison page skill | Claim + entity rules | Yes | Required |
| Full workflow | Stage orchestration | Yes | End-to-end |
If you already use marketing agent skills language, this pattern is the BOFU specialization for pairwise pages. Extraction patterns from turn marketing expertise into ai skill apply: pull pause points from how your best editor reviews comparisons.
Comparison page skill anatomy
The Comparison page skill anatomy breaks the file into four blocks every production skill should expose.
What inputs does a comparison page skill require?
Required keys typically include `competitor_a`, `competitor_b`, `icp_segment`, `risk_tier`, `research_packet_id`, and `citation_strictness`. Optional keys cover locale, industry, and schema variant. Invalid inputs fail at intake, same pattern as API validation.
Which rules prevent comparison page failures?
Rules read like linter messages. Each rule ties to a failure code agents return when checks fail. Rules must be falsifiable: “sound professional” fails; “every pricing mention includes a URL from the research packet” passes.
What examples should golden tests include?
Include one pass example and one fail example per high-risk rule. Show boundary cases: renamed product tier, deprecated feature, co-equal positioning requirement.
How do eval hooks tie skills to CI?
Declare rubric weights and minimum pass score. Link to golden inputs stored beside the skill tag. Eval hooks integrate with marketing skill evaluation practice.
| Block | Owner | Updates when |
|---|---|---|
| Inputs schema | GTM engineer | New competitor vertical |
| Rules | Content ops | Legal feedback |
| Examples | Senior editor | Rebrand |
| Eval hooks | Marketing ops | QA threshold change |
Rules that prevent comparison page failures
These rules belong in almost every production comparison skill. Adapt weights by risk tier.
- Entity lock: Product names and tiers must match the research packet JSON; fail on fuzzy match.
- Citation lock: Any numeric claim or pricing statement requires a packet URL; fail if missing.
- Neutrality: No superlatives unless packet includes cited proof; flag attack lines.
- Structure: Require summary table + dimension rows aligned to schema.org Product fields where CMS supports it.
- Links: Internal links only from approved live slug list; fail on draft paths.
- Freshness: Reject packet chunks older than configured SLA for pricing facts.
Align high-risk tiers with marketing agent guardrails: human approval before publish when score falls below threshold.
Worked failure: tier name drift
A draft used last year’s tier label while the research packet had been refreshed. Entity lock failed before CMS push. Fix: bump skill patch version, add counterexample, rerun golden set. Without the skill, the error would have shipped as “fluent” prose.
Sample input payload
Keep inputs small and explicit so workflows pass consistent keys.
| Field | Example | Purpose |
|---|---|---|
| competitor_a | Acme CRM | Left column entity |
| competitor_b | Beta CRM | Right column entity |
| icp_segment | Mid-market PLG | Tone and depth |
| risk_tier | high | Gate strictness |
| research_packet_id | rp_2026_07_30 | Single source of truth |
| citation_strictness | strict | Fail on missing URLs |
Validate the payload against JSON schema at workflow intake. Reject runs early when IDs point to stale packets.
Maintenance checklist for ops
Review the skill on a quarterly cadence or when a competitor rebrands.
- Confirm golden set still reflects live product tiers.
- Update counterexamples when legal adds a new banned claim class.
- Bump patch version for rule text fixes; minor for new optional inputs; major when pass or fail behavior changes.
- Log changelog entries next to the tag marketing ops approves.
- Re-run golden tests before changing the production pin.
Document owners in the same registry you use for workflow manifests so incident review does not hunt Slack for who approved v1.4.
Golden tests for comparison skills
Build 12, 20 golden pairs: mix easy passes, entity traps, missing citations, and neutrality violations. Score with the rubric declared in eval hooks. Hold out 20% for regression on every promotion.
| Golden case type | Expected result |
|---|---|
| Valid pair + full packet | Pass |
| Wrong tier string | Fail entity lock |
| Claim without URL | Fail citation lock |
| Aggressive competitor line | Fail neutrality |
| Link to unpublished slug | Fail link rule |
Block promotion when pass rate drops more than five points without owner sign-off. Document results next to the skill tag so incident review starts from data.
Wire the skill into the comparison pipeline
In the seven-stage comparison workflow, invoke the skill at draft and again at QA. Draft renders prose from schema; QA re-validates rules against the same packet ID.
``` Research packet → Schema JSON → comparison page skill (draft) → QA skill → Publish ```
Pin skill versions in the workflow manifest. Co-promote with eval rubric changes. Never edit the production skill tag without golden tests, same promotion ladder as other reusable marketing workflows artifacts.
Start with one competitor pair. Extract rules from how your editor marked up the last manual comparison. Codify, parameterize, evaluate, then tag v1.0.0. Scale triggers only after golden pass rates stabilize.
Most teams do not need a longer prompt, they need a versioned skill that agents call with stable inputs and explicit fail codes.
When comparison volume grows, judgment must compound: rules, examples, and eval travel with the skill instead of living in chat. Metaflow treats skills as first-class assets you pin, test, and attach to flows, so the standard you harden on pair one is what pair fifty inherits.
Frequently Asked Questions About Building a Comparison Page Skill
What is a comparison page skill?
It is a versioned instruction file with inputs, rules, examples, and eval hooks that agents use when drafting or checking pairwise comparison copy from structured research. It encodes editorial and legal judgment separately from layout templates.
How is a comparison page skill different from a prompt?
Prompts live in sessions and drift. The skill file is parameterized, testable, and tagged for promotion. Workflows call it by name and version. Metaflow stores skills outside chat so draft and QA stages invoke the same rule set by tag.
What rules belong in a comparison page skill?
Entity lock, citation lock, neutrality checks, required structure, internal link allowlists, and freshness SLAs for cost claims. High-risk tiers add human approval gates. Metaflow teams often map each rule to a visible fail code in QA output for faster review.
How do you test a comparison page skill?
Run golden competitor pairs with expected pass or fail labels. Regression-test on every skill promotion. Metaflow eval flows can run those golden sets against a candidate tag before you change the production pin.
Where does a comparison page skill sit in a workflow?
It typically runs at draft and QA in the comparison pipeline, after schema build and before CMS publish. Workflows pass research packet IDs and risk tier parameters. Metaflow flows wire the skill as a stage with logged inputs so debug traces show which rule failed.
