Agent Skills Reference

How to read this page

Every skill depends on the CompetLab MCP server being configured — that’s how it reaches your projects, competitors, and dimensions (see Setup). Six of the seven also need a project with competitors being monitored; competlab-site-audit needs neither a project nor configured competitors, only an API key. Skills trigger on plain-language requests; the Ask examples are phrasings each skill advertises, not fixed commands. The Reads line names the MCP tools a skill is allowed to call — they’re documented in the tool reference.

Which skill?

competlab-ai-visibility

Which companies do AI models recommend in this category — and is the customer one of them?

Reads the AI Visibility market map across the five engines — ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews — and gives the customer’s verdict before it shows a figure: Core, Too early to tell, or Rarely recommended. A small, stable set of companies is named again and again, and the rest are named seldom; the skill states which the customer is in first, with its presence, its range, and the count it rests on. Presence, ranges, and per-engine splits are the mechanism that decides membership, not the answer, and the blended AI Visibility Score is never what it leads with.

It checks the project’s prompt-market state before using the map: if the prompts return registries rather than vendors, it says so and does not lead with the map. Order is by presence — how often a brand is named. Two brands whose ranges overlap are tied, and the skill says tied. Google AI Overviews names companies in prose and ranks nothing.

• Reads: list_projects, get_project, list_competitors, get_ai_visibility_dashboard, get_ai_visibility_trend, get_ai_visibility_history, get_ai_visibility_check_detail.
• Ask: “who do AI models recommend in my space” · “are we in the core” · “AI visibility report” · “what does ChatGPT say about us” · “GEO analysis” · “market map”.
• Not for: traditional SEO or Google rankings, or which pages the models read to decide — that’s competlab-ai-sources.

competlab-ai-sources

Which pages do the engines read before answering, and who is named on them?

AI Visibility answers who is recommended; this skill opens up the next question — why them and not us. For the two engines that hand back the pages they retrieved while answering — Perplexity and Google AI Overviews — it reports, per engine, which companies each answer named, which pages the engine pulled, and which of those pages name the customer’s competitors and not the customer. That last list, split into approachable third-party hosts and rivals’ own sites, is the work list.

Every figure is a count with its universe, per engine, never pooled — and retrieved, never cited, because the engines do not disclose which pages they leaned on. Two limits travel with every claim: training data often outweighs what the models read, and a source can be read by every engine and belong to a company none of them recommend. It shows what the models read; it does not promise that working the list produces recommendations.

• Reads: list_projects, list_competitors, get_ai_sources_dashboard, get_ai_sources_history, get_ai_sources_check_detail.
• Ask: “why are they recommended and not us” · “what does AI read about my market” · “which pages decide the AI answers” · “where do I need to appear” · “AI sources report”.
• Not for: whether AI recommends you at all — that’s competlab-ai-visibility.

competlab-briefing

What changed, what it means, what to do — at pulse, landscape, or dimension depth.

Reads the Strategic Briefing the platform already wrote — its own synthesis across 14 analysis areas, in numbered editions that persist — at the depth the question needs. A pulse reads the executive digest; “what to do about it” leads with the top moves the digest names (every move in the briefing lands on your Strategic Tickets board — as a new ticket, most important first, or on the ticket already there for that work), then reads the edition’s tickets; a board review or full landscape adds the competitor standings and the two or three deep sections the digest flags; one dimension in depth reads that section alone. It compares editions to say what moved since last time, and reads the alerts that fired since the edition’s date.

It checks the briefing’s status first. While a run is in progress, or when the latest run produced no edition, it reads the most recent completed edition rather than reporting “no briefing”. It reads the briefing; it does not rebuild it.

• Reads: list_projects, get_project, list_competitors, get_briefing, get_briefing_history, get_briefing_edition, list_tickets, list_alerts.
• Ask: “briefing” · “CMO report” · “strategic briefing” · “competitive update” · “what changed with competitors” · “catch me up” · “competitive landscape” · “quarterly review” · “full competitive analysis” · “board-level CI”.
• Not for: a single competitor (competlab-competitor-dive) or a sales card (competlab-battlecard).

competlab-competitor-dive

What has this competitor actually done, what did the platform measure, and what does it change for us?

One rival, every dimension filtered to them — AI Visibility, AI Sources, Positioning, Pricing Intelligence, Content Intelligence, Tech & Trust Profile — plus the Strategic Briefing’s own researched read on that rival: funding, hiring, product launches, reliability, customer voice, developer ecosystem, agent adoption. The output is a dossier a founder, PMM, or sales leader acts on, closing with a SWOT where every item names the measurement and the run it came from.

Web search is the exception, not a step: it fills a gap no dimension and no briefing section covers, and anything it returns is verified with a fetch before it reaches the dossier.

• Reads: list_projects, get_project, list_competitors, get_competitor, get_briefing, get_briefing_history, get_briefing_edition, the positioning, pricing, content, tech & trust, AI Visibility, and AI Sources dashboards, positioning and pricing history, get_ai_visibility_trend, get_ai_visibility_history, get_ai_visibility_check_detail, get_content_changelog, and fetch_url; web search for gaps.
• Ask: “analyze Rival Inc” · “deep dive on Rival Inc” · “competitor dossier” · “competitor profile” · “SWOT analysis for Rival Inc” · “tell me everything about Rival Inc”.
• Not for: a card to read on a sales call (competlab-battlecard), or the whole market at once (competlab-briefing).
• Needs: a project where this competitor is monitored. A rival that isn’t monitored gets an offer to add them, not a dossier built from unverified web research.

competlab-battlecard

A 60-second sales reference for a live call.

Turns CompetLab monitoring data into a sales-ready battlecard against one competitor: at-a-glance comparison, why we win, where they are genuinely strong, objection handling, feature matrix, killer facts, and landmines. Built for a rep scanning it in sixty seconds before a call — scannable, spoken, sourced, honest, and dated. Every figure carries the count it came from, so it survives the prospect’s follow-up question, and the competitor’s real strengths are on the card in their own section.

It verifies the competitor’s price with a live fetch before a rep quotes it, and reads the briefing’s customer-voice section for what real users say rather than researching it again.

• Reads: list_projects, get_project, list_competitors, the pricing, positioning, tech & trust, content, AI Visibility, and AI Sources dashboards, get_ai_visibility_trend, get_briefing, and fetch_url; web search for an objection nothing else answers.
• Ask: “create a battlecard” · “sales battlecard for Rival Inc” · “why us vs Rival Inc” · “how to beat Rival Inc” · “objection handling for Rival Inc” · “sales cheat sheet” · “competitive one-pager”.
• Not for: the full dossier behind it — that’s competlab-competitor-dive.
• Needs: a project where this competitor is monitored.

competlab-site-audit

Can an automated system extract this company’s signals from its site?

Audits a site the way an automated reader meets it: which AI crawlers robots.txt admits, what the sitemap exposes, which tech and trust signals a scanner can actually extract, the agent-adoption level with the scored checks that failed, and whether a claimed MCP server answers the protocol or only answers a browser — a JSON-RPC POST decides that, not a 200 on GET. Findings are ordered by what a fix costs against what leaving it costs.

Two modes. On any public domain it runs the free scan tools with no project and no competitors configured — a free-trial key is enough. With a project it audits the customer’s own site and cross-checks it against what the monitored dimensions could and could not read from it, separating “the page is there and the extractor could not use it” from “nothing can read this site”. Blocking model-training crawlers is reported as a neutral content decision, never a gap.

• Reads: check_ai_crawlers, check_sitemap, fetch_url, and the three scan pairs — start_agent_adoption_scan / get_agent_adoption_scan, start_tech_stack_scan / get_tech_stack_scan, start_trust_signals_scan / get_trust_signals_scan. With a project: list_projects, list_competitors, and the tech & trust, positioning, pricing, and content dashboards. A shell command for the MCP verification.
• Ask: “audit my site” · “audit example.com” · “can AI crawlers read us” · “are we blocking ChatGPT” · “is our site set up for AI agents” · “trust signals check” · “tech stack scan” · “sitemap coverage” · “is that MCP server real”.
• Not for: what the models say about the brand (competlab-ai-visibility) or which pages they read while answering (competlab-ai-sources).
• Needs: a CompetLab API key — the free tools are still tools on the authenticated MCP server. No project, no competitors configured.

competlab-monitoring-setup

Am I monitoring the right competitors, asking the right questions, and running at the right cadence?

Reviews how a project is configured and recommends what to change. It checks the AI Visibility prompts first, because if they don’t describe the market the customer competes in, every other reading inherits the problem. It reads the Strategic Briefing’s own promotion suggestions and the dashboard’s list of brands core to the market but not on the roster, and frames every roster change as a swap — the brand in, the brand out, and the evidence on both sides — because a project monitors a limited number of competitors. Schedules are reported exactly as the API returns them.

It carries no write tools and changes nothing. The roster, the prompts, and the schedules are edited in the app by a person.

• Reads: list_projects, get_project, list_competitors, list_schedules, get_ai_visibility_dashboard, get_ai_visibility_trend, get_briefing.
• Ask: “am I tracking the right competitors” · “should I add Rival Inc to monitoring” · “are my prompts right” · “review my monitoring setup” · “how often does this run” · “who should I drop”.
• Not for: reading what the monitoring found — that’s the per-dimension skills.

The full suite

competlab-ai-visibility       which companies AI models recommend, and are we one of them
competlab-ai-sources          the pages the engines read, and who is named on them
competlab-briefing            what changed, what it means, what to do — pulse to landscape
competlab-competitor-dive     one rival, every dimension, plus the briefing's read on them
competlab-battlecard          a 60-second sales card for a live call
competlab-site-audit          what a machine can read on a site — no project needed
competlab-monitoring-setup    the right competitors, the right prompts, the schedules

What every skill carries

Each skill’s folder holds two reference files the agent reads before it writes: the platform facts — six dimensions, five engines, the tool map, the endpoint — and the reporting rules the platform enforces on itself. null means not measured, never zero — except a rank, which is null on a brand named in no answer: not named in any answer, a measured absence. A page of a long list is not the whole list. Counts, never rates. Per engine, never pooled. Retrieved, never cited. Overlapping ranges are not ordered. Model prose is the model’s, reported as what that model said and never as fact. The rules are what stop an agent from rounding, pooling, and ordering competitive data into something confident and wrong.

The data these skills read comes through the MCP server — see its tool reference for the underlying tools, and Setup to install the plugin. The suite is released on GitHub; this page describes v3.3.2 .