Radarkit

View saved version →

---
name: competitor-share-of-voice
description: Show which competitors and which source websites appear in the AI answers a Radarkit project tracks, as share of voice and citation counts, and where the brand ranks against them. Use when the user asks who they are losing to in AI answers, which competitors show up most, which sites AI models cite, or where they should try to get mentioned.
---

Use this skill when the user asks about competitors, share of voice, citations, sources, or rankings in AI answers.

1. Call `list_projects` and pick the project (ask if there are several and none was named).
2. Default the window to the last 30 days. Row-level tools accept at most 31 days per call;
   aggregates accept more.
3. For competitors, call `get_competitors`. It returns the brand's own share and every
   competitor's share of mentions with an average position.
4. For the websites AI answers rely on, call `top_source_domains`. `cited` tells how often a
   domain was actually cited rather than only consulted.
5. For a specific competitor or domain, call `list_rankings` with `domain` to see the exact
   answers and positions, or `list_sources` with `domain` for the exact pages.
6. Report the brand's share first, then the top five competitors with their share, then the
   top five source domains with citation counts. Point out the biggest gap and one concrete
   opportunity, for example a domain that is cited often where the brand has no page.

Do not present share of voice as a percentage of all AI traffic; it is the share of mentions
inside the answers this project tracks. Do not merge domains that differ only by subdomain.
Do not page through more than two pages of rows without asking; for bulk analysis hand off to
the export-answers skill.

Referenced files: 1

View saved version →

---
name: export-answers
description: Export the AI answers, sources, competitors, rankings or search queries a Radarkit project tracks as a downloadable CSV or NDJSON file for a date range, with every answer rendered as Markdown. Use when the user wants a spreadsheet, a file, a bulk download, "all answers", or more than a few pages of data.
---

Use this skill when the user asks for a file, a spreadsheet, a full export, or more rows than a couple of pages would return.

1. Call `list_projects` and pick the project (ask if there are several and none was named).
2. Decide the export `type` from the request: `responses` for the answers themselves,
   `sources` for cited pages, `competitors`, `rankings`, or `query_fanout` for the searches the
   models ran. Default to `responses`. Default `format` to `csv`; use `ndjson` only if asked.
3. Set the window. Exports accept up to 365 days. Narrow to specific topics with `topic_ids`
   (from `list_topics`) or specific prompts with `prompt_ids` when the user names them.
4. Call `estimate_export` first. It is free. Tell the user the row count and how many Export
   rows they have left, and ask for a go-ahead before creating the export. Never create an
   export without that confirmation.
5. Call `create_export`. It spends Export rows equal to the row count. Then call `get_export`
   with the returned id every few seconds until `status` is `ready` (usually under a minute).
   `get_export` is free.
6. Hand the user `download_url` and the file name. The link is valid for 30 minutes; the file
   stays for 7 days. Say what the columns are: prompt, topic, model, date, brand mentioned,
   position, sentiment, competitors, sources cited and the answer as Markdown.

Do not page through `list_responses` to build a spreadsheet yourself. Do not create an export
when the estimate says the account lacks Export rows; tell the user how many are needed.
If `get_export` reports `failed`, tell the user the rows were refunded and offer to retry.

Referenced files: 1

View saved version →

---
name: manage-prompts
description: Add the questions (prompts) a Radarkit project tracks, group them into topics, and re-ask a prompt to every AI model right now. Use when the user wants to track new questions, organise prompts into topics, or get a fresh answer for a prompt immediately instead of waiting for the next scheduled run.
---

Use this skill when the user wants to add, group, or immediately re-run tracked prompts. It needs a key or connection with the Write permission; if the tools are not listed, say so.

1. Call `list_projects` and pick the project (ask if there are several and none was named).
2. Call `list_topics`. Reuse an existing topic when its name fits. Create one with `add_topic`
   only when nothing fits; the server refuses duplicates and returns the existing topic id.
3. Turn the user's request into prompt texts written the way a real person would ask an AI
   assistant: one question each, under 255 characters, no numbering. Show the list and ask
   for confirmation before adding.
4. Call `add_prompts` once with the topic id and up to 50 prompts. Report what was added and
   what was skipped as a duplicate. New prompts are answered on the project's next scheduled
   run; say when that is (`next_run_at` from `get_project`).
5. Only when the user explicitly asks for an answer now, call `run_prompt_now` for that one
   prompt. It spends one realtime credit from the project owner's balance, so confirm first.
   Afterwards, the fresh answers appear in `list_responses` with `realtime: true` within a few
   minutes; poll `list_prompts` until the prompt's model states read `completed`.

Do not add prompts the user did not confirm. Do not run a prompt now on the user's behalf as
a side effect of adding it. Do not run the same prompt again while it is still processing;
the server answers PROMPT_PROCESSING, wait for it to finish.

Referenced files: 1

View saved version →

---
name: visibility-report
description: Report how visible a brand is in AI answers (ChatGPT, Gemini, Perplexity, Copilot, Google AI Mode, AI Overviews) for a Radarkit project over a date range, broken down by model and by day, with the sentiment and position of the mentions. Use when the user asks how their brand is doing, how visible they are, whether visibility went up or down, or which AI model mentions them least.
---

Use this skill when the user asks about their brand's visibility, mentions, or trend in AI answers.

1. Call `list_projects`. If the account has one project, use it. If it has several and the user
   did not name one, ask which project before spending credits.
2. Work out the window. Default to the last 30 days ending today. Accept "last week", "this
   month", "since 1 August" and convert to `from` and `to` as YYYY-MM-DD. Visibility accepts up
   to 366 days in one call.
3. Call `get_visibility` with the project id and the window. Read `summary` for the overall
   score, `by_model` for one row per AI model, and `series` for the day-by-day line.
4. If the user asks why, or which answers drove the number, call `get_sentiment` for the
   sentiment split and `list_responses` with `brand_mentioned: false` to show the answers that
   left the brand out. Keep `include_content` off unless the user wants the answer text.
5. Report: the overall visibility percentage, the best and worst model with their numbers,
   the direction of the trend (compare the first and last week of the series), and the
   sentiment split. Name the models the way Radarkit does: ChatGPT, Gemini, Perplexity,
   Copilot, Google AI Mode, AI Overviews. Mention how many API credits were spent.

Do not average numbers across models yourself; `summary` already does that. Do not describe
days that have no data as zero visibility; say there was no run that day. Do not call the
same tool twice for the same window.

Referenced files: 1