agentsclimarketplace
Sign in

Aeko prompt deep dive

Skill AEKO-Intelligence/aeko-plugin/skills/aeko-prompt-deep-dive

Citation source-analysis deep-dive on one tracked prompt. Pulls per-platform responses, citation arrays, and crawled source metadata, then helps the user understand which competitors win the prompt, which sources AI engines cite, and what those sources do that could be mirrored. This skill operationalizes AEKO's unique cited-source analysis value-prop — vanilla Claude cannot do it.From its SKILL.md

Install
npx -y skills add AEKO-Intelligence/aeko-plugin --skill aeko-prompt-deep-dive

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

9.1 KB, as published. Nobody here has run it

AEKO Prompt Deep-Dive

Walks a user through one tracked prompt's full citation footprint. Output: a clear picture of which pages / brands win this prompt across AI engines, plus one concrete action the user could take.

Marketer-facing output contract

Explain the prompt as a real customer question. Keep the report focused on: who AI mentions, which pages AI cites, what facts AI appears to reuse, and the one best action. Label fact reuse as inferred absorption, not measured fact.

Plain words, not jargon. Never surface "forensics" / 포렌식 in user-facing copy. In English say source analysis; in Korean say AI 답변 참고 출처 (the sources AI references in its answers). "Forensics" elsewhere in this doc is an internal label only.

Language: mirror the user's chat language for user-facing steps, summaries, questions, and next actions. Keep slash commands, IDs, file paths, citation URLs, schema keys, and tool names in English/ASCII.

Input

  • prompt-id (required) — $1. UUID of a tracked prompt (must have a UserPrompts row for the current user).
  • window (optional) — $2. latest (default), 7d, 30d, or 90d.

If prompt-id is missing, tell user to get one from /aeko-find-prompts-to-track or /aeko-action-center.

Step 1 — Fetch the source-analysis payload

Call aeko_get_tracked_prompt(prompt_id, window=<from $2 or "latest">).

On 404 → the user isn't tracking this prompt (or never did). Tell them and suggest /aeko-find-prompts-to-track to discover + track candidates.

On success, the payload includes:

  • prompt: text, language, country, industry, vertical, query_type, funnel_stage, context.
  • responses: array of per-platform responses (ai_platform, response_date, full_response, mention_count, citation_count, source_count, sentiment, observed_intent, mentions, citations[], citations_truncated).
  • Each citation: source_url, domain, source_type, mention_name, position_in_response, context_snippet, crawl (or null).
  • Each crawl (when non-null): extracted_text (may be truncated to 5000 chars, flagged), meta, json_ld[], source_analysis.

Step 2 — Summarize per-platform behavior

Print a header block:

# Deep-dive: <prompt text, truncated to 100 chars>
- ID:        <prompt_id>
- Window:    <window>
- Prompt context: <country> · <industry> · <query_type> · <context>

Then a per-platform summary table:

| Platform | Date | Mentions | Citations | Sentiment | Brand present |
|----------|------|----------|-----------|-----------|---------------|
| Claude   | ...  | 3        | 2         | neutral   | No            |
| GPT      | ...  | 2        | 2         | positive  | Yes (pos 3)   |
| Gemini   | ...  | 1        | 0         | —         | No            |

"Brand present" checks whether the user's brand appears in mentions (compare against the domain name, base URL host, domain keywords, and names surfaced in aeko_get_domain_info). If yes, show position if derivable from context_snippet.

Step 3 — Rank cited sources

Aggregate across all platforms in the window. For each unique domain + canonical_url pair:

  • total_citations = sum across responses.
  • avg_position = mean position_in_response (ignore nulls).
  • platforms = set of AI platforms that cited this source.

Sort by total_citations desc, then avg_position asc (lower = earlier in AI output = more prominent).

Print top 5-10:

## Top cited sources

1. **<domain>** · <source_type> · cited <N>× across {Claude, GPT} · avg pos 2.5
   URL: <source_url>
   Mentioned alongside: <brand names from context_snippet, comma-separated>
   → Why AI cites it: <one-line hypothesis based on crawl data below>

2. ...

Step 4 — Structural analysis of each top source

For each top source with non-null crawl, analyze:

  • JSON-LD types present (from crawl.json_ld[] — look at @type field). Call out FAQPage, Product, Review, AggregateRating, Article, BreadcrumbList.
  • Source analysis signals (from crawl.source_analysis): citability score if present, heading depth, structural patterns.
  • Content shape (from first ~500 chars of crawl.extracted_text): Q&A format, first-person review, comparison table, listicle, news article, etc.
  • Framework read — why does this source get cited? Name it in the plugin's AEO vocabulary (BLUF / PREP / Informational Gain / E-E-A-T — see skills/aeko-create-content/references/aeo-frameworks.md): does it lead with the answer (BLUF)? give self-contained Point·Reason·Example blocks (PREP)? carry lived, specific detail a generic page lacks (Informational Gain)? show experience in its FAQ (E-E-A-T)? This is what makes the takeaway actionable — it maps straight to a fix the executor skills apply.
  • Inferred absorption — compare response.full_response, citations[].context_snippet, and crawl.extracted_text for overlapping facts, phrasing, numbers, and claims. Label this as inference: "AI appears to use..." Never present it as a measured metric.

Print under each top source:

   Structure:  <content shape>
   JSON-LD:    <types present, or "none">
   Citability: <score if known, else "N/A">
   Wins via:   <framework(s) it exploits — e.g. "BLUF + E-E-A-T FAQ">
   AI appears to use: <inferred facts/phrasing, or "not enough evidence">
   Takeaway:   <one sentence: the framework gap to close on the user's page>

If crawl is null (source never crawled) and the URL is public, do a light WebFetch to get a rough structural read — mark the analysis as "live-fetched, not cached" so the user knows the difference.

Step 5 — Competitor callout

From the aggregated mentions field across responses, build a frequency table of brand names. Flag the user's own brand using domain info + the top 3-5 competitors by total mention count:

## Who's winning this prompt

- **<competitor>**: mentioned <N>×, cited <M>× — appears in {Claude, GPT}
- **<competitor>**: ...
- **<user's brand>**: mentioned <N>× (target: grow this)

If the user's brand is absent, say so plainly: "Your brand isn't surfacing for this prompt yet. That's the gap to close."

Step 6 — One concrete action

Based on the analysis, propose exactly ONE action the user could take. Name the framework gap behind it (from Step 4's "Wins via") so the action is unambiguous — e.g. "the cited winner leads with a BLUF answer; your PDP buries it under specs" or "winners show Informational Gain from real usage; yours reads generic." Pick from:

  • Mirror a winning PDP structure — "Three top cited sources are retailer PDPs with FAQPage JSON-LD. Your matching PDP <url> doesn't have FAQPage markup. Run /aeko-action-center <domain_id> pdp to see if there's a pending PDP action for this product."
  • Write content to match a winning format — "The top cited source is a Naver 블로그 first-person review. AEKO doesn't have an action item for this, but drafting your own review-style piece could help. Run /aeko-action-center <domain_id> content to see pending content items, or tell me to draft one standalone."
  • Close a technical gap — "Top cited sources all have Organization JSON-LD with sameAs pointing to Wikipedia/Wikidata. Your domain's JSON-LD is missing. Run /aeko-action-center <domain_id> technical for the matching fix."
  • Track a related prompt — "This prompt is awareness-stage. The brands winning it also win comparable consideration-stage prompts (e.g. '<related query>'). Run /aeko-find-prompts-to-track to add related prompts to your watchlist."

Pick the most specific, load-bearing action. If none fits, pick "run the visibility report for context" (/aeko-visibility-report <domain_id>).

Step 7 — Save the analysis

Write the full analysis markdown to ./aeko-artifacts/<domain_id>/prompt-deep-dives/<prompt_id>-<window>.md so the user can revisit without re-running. Tell them the path.

Error paths

  • 404 from aeko_get_tracked_prompt → suggest /aeko-find-prompts-to-track.
  • Zero responses in the window → tell user AEKO hasn't collected data for this prompt yet; suggest checking back in 1-3 days OR widening window to 90d.
  • All citations have null crawl → analysis falls back to WebFetch for top 3 sources only; cap to avoid flooding the skill.

What this skill never does

  • Never untracks the prompt or modifies tracking state (that's /aeko-find-prompts-to-track for adding, dashboard for removing).
  • Never fabricates citation data — if a source URL returns 404 on live fetch, note it and move on.
  • Never creates an action item — only suggests commands the user runs.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.