Menu

Browse all skills

Content agent skill

Content

take a keyword or topic through research, structure, and writing to a publish-ready draft, not just an outline. The rules live in [references/query-types.md](references/query-types.md) and [references/writing-and-titling.md](references/writing-and-titling.md). This skill is the process that applies them to a real topic. Evidence rules are in [references/evidence-standards.md](references/evidence-standards.md).

View raw SKILL.mdView package on GitHub

When to use this skill

Use when the user wants a finished article, or an existing article rewritten or optimized, for a single query or topic.

Do not use when:

  • The user only wants a plan. Use content-brief.
  • The topic is broad enough to need a hub with spokes. Use pillar-article.
  • The user wants a multi-asset, industry-specific package (pillar, supporting articles, social, email). Use industry-content-production.
  • The user only wants a standalone FAQ set. Use faq-generator.

Required inputs

target_query: ""          # keyword, prompt, or topic
brand:
  name: ""
  website: ""
  offer: ""               # what the CTA should lead to
market:
  country: ""
  language: ""
brief: ""                  # optional: content-brief output; if present, follow it
voice_sample: ""           # optional: URL or text of existing brand copy
existing_pages: []         # optional: URLs available for internal links
prior_findings: ""         # optional: competitor-audit / keyword-research output
sources: []                # optional: user-supplied facts, data, quotes

If target_query or brand is missing, ask. If brief exists, skip steps 1–2 except to check that the brief is still current.

Workflow

- [ ] 1. Query type, funnel stage, and scope classified
- [ ] 2. Live results and prior findings reviewed; angle chosen
- [ ] 3. Outline built (H1, sub-question H2s, formats, FAQ, link map)
- [ ] 4. Title tag and meta written
- [ ] 5. Draft written; every factual claim sourced or flagged
- [ ] 6. scripts/check-draft.mjs passes; manual checklist passes
- [ ] 7. Output assembled with source list and flagged claims
  1. Classify.
    • Assign the query type and funnel stage using references/query-types.md. This drives every structural choice. Do not guess. If the query is genuinely mixed, pick the dominant intent and say why.
    • Write one sentence stating what the page covers and one stating what it does not.
  2. Research.
    • Review the current live results for the exact target query: which formats win, what the strongest pages do well, and what is missing.
    • Reuse existing competitor-audit or keyword-research findings instead of re-researching. Record retrieval dates.
    • Choose an angle based on first-hand or non-commodity value the brand can actually provide, such as data, experience, or specifics. If no live search access is available, say so and label the research as assumption.
  3. Outline.
    • H1 that matches the query near verbatim.
    • H2s written as real sub-questions, each assigned a format that fits the query type (table, steps, list, grouped by period).
    • A table of contents if the piece is long.
    • A FAQ block of 3–6 adjacent long-tail questions the body does not answer.
    • An internal link map: pages to link out to, and pages that should link back here.
  4. Title and meta.
    • Apply the title pattern for the query type from references/query-types.md.
    • The title tag should be about 50–60 characters and aligned with the H1 in substance.
    • The meta description should be about 155 characters or fewer, preview the specific answer, and not repeat the title.
  5. Draft. Follow references/writing-and-titling.md:
    • Open with the answer. Keep paragraphs to 2–4 sentences.
    • Use specific numbers and named entities instead of vague claims. Use semantic variants naturally, with no keyword quota.
    • Place internal links where the entity is mentioned, not grouped at the end.
    • End with one CTA that matches the brand's real funnel.
    • Every statistic, price, date, or named fact must trace to a source in the source list (see references/evidence-standards.md). If you cannot source a claim, mark it inline as [NEEDS SOURCE]. Never invent a number to fill a gap.
  6. Check.

    • Save the draft with title and meta_description frontmatter and run:

      node scripts/check-draft.mjs draft.md

    • Fix every fail. Review each warn.

    • Then confirm the manual items the script cannot judge:

      • each H2 answers a real sub-question
      • the format matches the query type
      • no keyword stuffing
      • the CTA matches the funnel

    • Repeat until both pass.

  7. Output using the template below.
    • If any [NEEDS SOURCE] markers remain, the status is needs_trust_qa, not publish_ready.
    • Hand off to trust-qa.

Output template

---
title: "<title tag>"
meta_description: "<meta>"
status: publish_ready | needs_trust_qa
query_type: <type>
funnel_stage: <stage>
---

# <H1>
<body … FAQ … CTA>

---
## Internal link map
| Anchor | Target URL | Direction (out/in) |

## Sources
| # | Claim | Source URL | Accessed | Confidence |

## Flagged claims for trust-qa
| Claim | Reason (unsourced / outdated / conflicting) |

## Check results
<check-draft.mjs JSON summary>

Approval gates

  • Do not publish or push to a CMS. Deliver the draft only, unless the user explicitly asks and grants access.
  • Get approval before naming competitors negatively, quoting real people, or making legal, medical, or financial claims.

Stop conditions

  • The query's intent cannot be determined and the user is unavailable to clarify.
  • The topic requires regulated advice (medical, legal, financial) and there is no qualified source or reviewer.
  • The core claim of the article cannot be sourced. Report this instead of drafting around it.

Anti-patterns

  • Writing from general knowledge without checking current results.
  • Answering a cost query in loose prose instead of a cost table, or making any other format mismatch.
  • Forcing a fixed count of bolded keywords, or stuffing variants.
  • Inventing statistics, quotes, or sources to make prose feel concrete.
  • Adding FAQPage or HowTo schema for rich results. Neither shows as a rich result any more (see references/evidence-standards.md).
  • Rewriting content "for AI" with artificial chunking or llms.txt-style tricks. Google says these are unnecessary.
  • Skipping the internal link map and leaving the page orphaned.

Completion checklist

  • Query type, funnel stage and scope stated
  • Live results reviewed (or labeled assumption). Angle based on real value the brand provides
  • Outline has sub-question H2s, assigned formats, FAQ and link map
  • Title and meta follow the rules. H1 aligned
  • Every factual claim is sourced or marked [NEEDS SOURCE]
  • check-draft.mjs passes with no fail. Manual checks done
  • Status set honestly. Flagged claims handed to trust-qa