Magister Marketing

View saved version →

---
name: magister-ai-visibility
description: Track and report how a brand appears in AI-generated answers — ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews — using Magister AI-visibility prompt sets. Use when the user asks about AI visibility, GEO, generative engine optimization, whether AI assistants mention or recommend their brand, or wants to run or read a visibility report.
---

# Magister AI visibility (GEO)

An **AI-visibility set** is a group of tracked prompts — the questions a brand's
buyers actually ask. Magister runs each prompt against several AI answer engines
and scores how the brand appears. Each execution of a set is a **round**, and
comparing rounds over time is the point of the product.

Resolve `project_id` from `list_projects` first; never guess one.

## Reading existing visibility

1. `list_ai_visibility_sets` — the project's tracked sets.
2. `get_ai_visibility_set` — one set's prompts, rounds, and latest scores.
3. `render_ai_visibility_report` — a readable score summary plus an interactive
   report with the trend across rounds, scores by AI model, and per-prompt
   detail.

Prefer `render_ai_visibility_report` when the user wants to *see* how they are
doing. Call it once; it needs no polling and performs no network fetches of its
own.

## Rounds that are not comparable are deliberately excluded

The trend only charts rounds measured the same way. A round that was scored
against a different set of models, or that failed to measure enough of its
prompts, is **left out of the trend on purpose** rather than averaged in.

So the chart can legitimately show fewer points than the raw round count. That
is correct behavior, not missing data. Do not describe it as an error, and do
not try to reconstruct the excluded rounds.

## Running a new round costs credits

`run_ai_visibility_set`, `create_ai_visibility_set`, and
`suggest_ai_visibility_prompts` all **spend the organization's credits**.

Before starting one, tell the user it is a metered run. If they only want to
know their current standing, read the existing set instead — reads are free.
`get_credit_balance` reports the balance and whether metered tools are currently
blocked.

## Rounds are asynchronous

`run_ai_visibility_set` starts a round; it does not return scores. Poll
`get_ai_visibility_set` until the round reaches a terminal state.

Never start a second round because the first has not finished — that spends
credits again for the same answer. Rounds take minutes, not seconds; tell the
user it is running rather than waiting silently.

## Creating a new set

Use `suggest_ai_visibility_prompts` to generate site-grounded prompt candidates,
show them to the user, and create the set only from prompts they accept.
`create_ai_visibility_set` can start the set's first round immediately, so say
that before calling it.

Write prompts the way a buyer would ask them — a category question the brand
should win, not the brand's own name alone.

## Do not infer

- Never state a visibility score, a change, or a model's answer the tools did
  not return.
- Never claim a brand is or is not mentioned by an AI assistant without a round
  that measured it.
- If a tool reports the organization is out of credits, say plainly that they
  need to upgrade, and stop. Do not retry and do not reach for another metered
  tool.

View saved version →

---
name: magister-marketing-plan
description: Work a brand's Magister marketing plan — find the highest-priority task, read its brief, record progress, add or edit tasks, and change roadmap strategy. Use when the user asks what to work on next, what their marketing plan says, how a task is going, or wants to add, update, complete, or reprioritize marketing work.
---

# Working a Magister marketing plan

A Magister **project** is one brand or client. Each project has one live
**marketing plan** built from that brand's audit evidence, organized into moves
across acquisition, activation, retention, referral, and revenue, and broken
into **plan items** — the individual pieces of work.

## Always resolve the project first

Every plan tool needs a `project_id`. Get it from `list_projects`. Never guess
one, never reuse an id from an earlier conversation, and never construct one.

If `list_projects` returns exactly one project, use it without asking. If it
returns several, ask which brand the user means before doing anything else.

## Answering "what should I work on?"

Call `get_next_plan_item`. It returns the single most actionable item **plus a
self-contained `brief_markdown`** that already explains the work, the reasoning,
and the acceptance criteria.

Present that brief. Do not re-derive the task from the plan, do not summarize it
into something shorter unless asked, and do not invent steps it does not
contain. If the user wants the wider picture instead, `get_marketing_plan`
returns the whole plan with its sections and items, and `get_plan_item` returns
one item in full including its recent timeline.

## Recording progress

- `update_plan_item_status` — mark an item running, completed, reopened,
  blocked, or skipped. Status changes are reversible, so no confirmation is
  needed beyond the user's request.
- `add_plan_item_note` — append a progress note to an item's timeline. Use this
  for findings and context that should survive the conversation.

Record status honestly. If work partly succeeded, say so in a note rather than
marking the item completed.

## Changing the plan itself

- `create_plan_item` adds a new item; `update_plan_item` edits an existing one's
  title, description, task instruction, expected impact, phase, due date, or
  priority. Both are ordinary edits.
- `set_plan_preferences` and `recompile_marketing_plan` are different in kind.
  **A recompile replaces the live plan's sections and items with a newly
  generated version.** Any hand-authored item the user added can disappear from
  the active plan.

Before calling either one, tell the user plainly that the current plan will be
regenerated and wait for them to confirm. `recompile_marketing_plan` also
**spends the organization's credits**, so say that too.

Changing only notes or non-roadmap preferences does not trigger a recompile.
When `set_plan_preferences` does start one, it returns a compile run — poll it
with `get_compile_run` rather than calling recompile again.

## Background work

A recompile runs in the background. Poll `get_compile_run` until it reports a
terminal state. Never start a second recompile because the first has not
finished, and never report a new plan before the run completes.

## Useful context tools

- `get_brand` — the audited brand profile: positioning, audience, voice, and
  messaging pillars. Read this before drafting anything in the brand's voice.
- `get_audit_health` — the overall health score and per-channel breakdown from
  the last completed audit. It needs the audit plan id that `get_marketing_plan`
  supplies.
- `list_recent_activity` and `list_agent_artifacts` — what has recently been
  worked on and shipped for the project.

## Do not infer

- Never state a score, status, due date, or priority the tools did not return.
- Never claim an item is done because you drafted something; only a status write
  makes that true.
- If a tool reports the organization is out of credits, tell the user they need
  to upgrade and stop. Do not retry it and do not substitute another metered
  tool.

View saved version →

---
name: magister-paid-ads
description: Review paid advertising performance and create new campaigns safely through Magister — Google Ads, Meta Ads, and TikTok Ads reporting, plus creating Meta or TikTok campaigns that are always paused. Use when the user asks how their ads are performing, what campaigns are running, wants to create or launch a campaign, or wants to stop ad spend.
---

# Paid advertising through Magister

This connector can **read** advertising performance across the user's connected
platforms and **create new Meta or TikTok campaigns in a paused state**. It
cannot start ad spend. Be precise about that boundary — it is the single most
important thing to get right here, because the user's money is involved.

Resolve `project_id` from `list_projects` first; never guess one.

## Reading performance

- `list_social_accounts` — which advertising and social accounts are connected.
- `list_ad_campaigns` — campaigns on those accounts.
- `get_ad_analytics` — performance metrics for one advertisement.

These read from the connected ad platforms and change nothing.

## Creating a campaign

`create_ad_campaign` supports **Meta and TikTok only**, and every campaign it
creates is **paused at creation** and verified paused before the call returns.

Two things to handle before calling it:

1. **A connected account can front more than one ad account.** When it does, the
   tool asks which one to bill. That is expected behavior, not an error — put
   the choice to the user rather than picking one. Billing the wrong ad account
   is a real financial mistake.
2. **Confirm the budget with the user** in the same message you confirm the
   platform and the audience.

If the user asks for LinkedIn, X, Pinterest, Google, or any other platform, say
directly that campaign creation supports only Meta and TikTok here, and that
reporting is still available for their other connected platforms. Do not
substitute a different platform to satisfy the request.

## This connector cannot launch a campaign

There is **no activation tool**. A campaign created here stays paused until a
person activates it in the Magister app or in the ad platform itself.

Say so plainly when you create one. Never imply the campaign is live, never
promise it will start on its own, and never describe a paused campaign as
"launched" or "running."

## Stopping spend

`pause_ad_campaign` stops further spend on a running campaign. It is reversible
by activating the campaign again in the ad platform, so it is safe to use when
the user asks to stop something. If a user is worried about spend, pausing is
the correct immediate action.

## Do not infer

- Never state spend, impressions, conversions, or ROAS the tools did not return.
- Never estimate what a campaign "should" cost or perform and present it as
  measured data.
- Never claim a campaign is active. Nothing on this surface can make it active.

View saved version →

---
name: magister-social-publishing
description: Draft, schedule, and publish social posts through Magister, and read post and follower analytics. Use when the user wants to write a social post in their brand voice, schedule content, publish a draft, see how posts performed, or find the best time to post.
---

# Social posts through Magister

Drafting and scheduling are ordinary writes. **Publishing to a public account
requires a person's approval** and cannot be forced from this connector.

Resolve `project_id` from `list_projects` first; never guess one.

## Every post needs a real account

Call `list_social_accounts` before creating a post. Each platform you post to
needs its **`account_id` from that list** — supplying a platform name alone is
rejected. Do not invent an account id and do not carry one over from a previous
conversation.

If several accounts are connected for a platform, ask which one.

## Drafting

`create_social_post` creates a draft. Read `get_brand` first so the copy matches
the brand's actual voice, audience, and messaging pillars rather than a generic
tone.

`update_social_post` edits an unpublished post. Note that it **overwrites the
caption with no revision history**, so if the user may want the previous version,
show it to them before replacing it.

## Scheduling commits to a future public post

Supplying a schedule time is not the same as saving a draft. Once approved, the
post will publish publicly at that time, and **this connector has no tool to
unschedule or delete it afterwards** — that has to be done in the Magister app.

Say that before scheduling, and confirm the time and time zone.

## Publishing requires approval, and that is not a failure

`publish_social_post` does not publish on the call. It creates a permission
request describing the exact post, and a person must approve it in the Magister
app before anything goes live.

When the tool returns that it needs approval:

- Tell the user the post is waiting for their approval and give them the
  approval link the tool returned.
- **Do not call the tool again.** A second call does not publish it and only
  creates confusion; the request is already waiting.
- Do not describe the post as published, live, or sent.
- `get_action_approval` checks one request's status and
  `list_action_approvals` lists pending ones, if the user asks what is waiting.

## Reading performance

`list_social_posts` returns the project's posts and their status.
`get_social_analytics` returns per-post metrics, daily numbers, follower stats,
and best-time-to-post insights for the connected accounts.

Use the posting-time data when the user asks when to post, rather than offering
generic advice about optimal posting times.

## Do not infer

- Never state engagement, reach, or follower numbers the tools did not return.
- Never claim a post published without a completed publish.
- Never work around the approval step, even if the user asks you to skip it.