agentleFS
Sign inStart free
gomarble-aigomarble-ai/marketing-agent/skills/automate-with-agents/SKILL.md

automate-with-agents

Skill15 starsChanged 38 days agoMITView on GitHub

Use when the user wants GoMarble to keep doing a job on a schedule: a daily brief, anomaly or spend alerts, a weekly creative fatigue check, a search term cleanup every Monday, budget pacing, competitor monitoring, or a recurring report by email or Slack. Covers creating, listing, editing, pausing, running and deleting GoMarble agents with manage_agents, including write agents that propose changes for approval.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Call manage_agents with action: "list".

    line 27
  2. Very specific:

    Call manage_agents with action: "create", type: "prompt", promptTitle, prompt, scheduleType, cronExpression, timezone, deliveryMethods and …

    line 45
  3. Distinctive:

    Never create, update, activate or delete an agent the user didn't ask for.

    line 98
  4. Very specific:

    Never turn on write access without the user saying so, and never set writeMode: "auto_apply" from here, even if asked (see Read or write …

    line 99

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Automate with agents

GoMarble agents run a standing job on a schedule in GoMarble's cloud, whether or not the chat app is open. Each run investigates what changed and delivers the result by email or Slack. An agent can be read-only (watch and report) or have write access (propose changes, applied only after approval by default).

"Agent" and "schedule" mean the same thing. Everything goes through one tool: manage_agents.

When to use

  • "Every morning, tell me how yesterday went across my accounts."
  • "Alert me when ROAS drops below 2 or spend spikes."
  • "Check my ads for creative fatigue every week."
  • "Clean up search terms every Monday and propose the negatives."
  • "Send the team a performance report every Monday at 9."
  • "Pause my agents", "what agents do I have?", "run the daily brief now".

For a one-off analysis, don't create an agent. Answer directly, then offer to put it on a schedule.

Workflow

1. See what already exists

Call manage_agents with action: "list". If an existing agent already covers the job, such as a daily brief, offer to update it rather than creating a duplicate.

2. Agree on the job

Settle these in one short exchange. Don't ask about anything the user already told you.

SettingHow to decide
The jobWhat to watch, which thresholds matter, what to do about it, and what the output should look like. Base it on a starter agent below when one fits.
AccountsUse account IDs exactly as the list tools return them (facebook_list_ad_accounts, google_ads_list_accounts, tiktok_list_ad_accounts, and so on). Account names in the prompt don't count as scope.
Cadencedaily, weekly or monthly, matched to the job. Don't over-notify. Convert the user's local time to a UTC cronExpression and pass their IANA timezone.
DeliveryThis is the user's choice, so ask. Email (gmail + emailRecipients) works for everyone. Slack (slack + slackChannels) needs Slack connected at apps.gomarble.ai/settings/apps (on a plan that includes it) and a channel the GoMarble bot can see. If they have no preference, propose email to their own address. Never default to Slack. Each method needs its destination in the same call, or the agent runs and delivers nothing.
When to notifynotificationMode: "always", or "conditional" with a notificationCondition such as "ROAS dropped below 1x".
Read or writeMonitoring, briefs, fatigue flags, competitor watch and alerts are read-only (the default). Only a job that should act on the account needs permissions: { accessLevel: "write", writeMode: "ask_approval" }. Write agents must include at least one Meta Ads or Google Ads account and need a GoMarble plan with write actions. From here, always use writeMode: "ask_approval": the agent prepares each change and it waits for someone to approve it in GoMarble. Never set writeMode: "auto_apply". If the user wants changes applied without approval, explain that a workspace owner or admin can set that up in the GoMarble web app, where it comes with guardrails such as budget-change caps.
GuardrailsFor write agents, offer guardrails: allowedActions (for example only status_toggle and budget_change) and maxBudgetChangePct (for example 20). These need the agent-governance plan feature. accountScopeStrict: true locks the agent to its accounts; ask before setting it.

3. Create it as a draft

Call manage_agents with action: "create", type: "prompt", promptTitle, prompt, scheduleType, cronExpression, timezone, deliveryMethods and destinations, plus the settings above.

A new prompt agent is a draft by default: saved and editable, but nothing runs or delivers until the user activates it. Tell the user it's a draft. Activate it with action: "resume" only after they confirm. Pass isDraft: false only when the user explicitly asked for it to go live right away.

Optional settings:

  • intelligenceLevel: instant, medium (default) or high for deeper reasoning.
  • memoryMode: auto (default), on or off. When on, the agent follows up on its own earlier recommendations.

4. Manage existing agents

User wants toCall manage_agents with
See their agentsaction: "list" (optionally type: "prompt" or "report")
Change the prompt, cadence, delivery or accountsaction: "update", scheduleId, and only the fields that change
Stop it for nowaction: "pause", scheduleId
Start it (or take a draft live)action: "resume", scheduleId
Run it once nowaction: "run", scheduleId
Remove itaction: "delete", scheduleId, only after the user confirms. Deleting can't be undone; pausing can.

Report agents (type: "report") re-run a report saved in the GoMarble app and need its reportId. To schedule a report built in this conversation, create a prompt agent whose prompt describes the report (see the build-reports skill).

Writing a good agent prompt

  • Name the accounts, the comparison windows (for example yesterday vs the trailing 7-day average, or 3-day vs 7-day), and the thresholds.
  • Say what counts as noise. "Default to HOLD" and "ignore entities changed in the last 48 hours" prevent over-reacting.
  • Say what to deliver: a maximum number of bullets, a table, or "only message me when something needs action".
  • For write agents, say exactly which changes it may propose and the limits (for example "budget changes of at most 20%").

Starter agents

These are based on GoMarble's own agent templates. Adapt them to the user's accounts and targets.

AgentPlatformsCadenceAccessThe job
CMO daily briefAll connectedDailyReadYesterday vs 7-day average for spend, ROAS and CPA per channel, plus GA4 sessions by channel and Klaviyo revenue. At most 5 bullets.
Performance analystMeta, GoogleDailyReadFlag the one or two account changes that deserve attention, with the likely driver and next step.
Creative fatigue monitorMeta, TikTokWeeklyReadFrequency up and CTR down more than 20% (first vs last 7 days of 28) = fatigued; 15–20% = creeping in.
Search term analyzerGoogle, MicrosoftWeeklyRead or writePareto the search terms driving 80% of spend and flag wasteful ones. With write access, propose negatives for approval.
Budget pacing / optimizerMeta, GoogleDailyRead or writeDefault to HOLD. Use 3-day and 7-day ROAS trends. With write access, propose budget moves of at most 20%.
Impression share trackerGoogleWeeklyReadLost IS (budget) vs lost IS (rank) on profitable campaigns: raise budget, improve rank, or do nothing.
Bids optimizerGoogle, MicrosoftWeeklyRead or writeAlign tCPA/tROAS targets with actual performance. Never change strategy under 30 conversions a month or move targets more than 20%.
Competitor radarAd libraryWeeklyReadNew and scaling competitor ads, grouped by angle, with the gaps worth testing next.
Monthly budget reallocation memoAll connectedMonthlyReadFour weeks of cross-channel efficiency trends and a one-page recommendation for next month's split.
CAC vs LTV trackerAds + Shopify + KlaviyoMonthlyReadBlended CAC, 90-day LTV per cohort and LTV:CAC by month.

Permissions

  • Team members need the Can build agents permission to create or change agents. If manage_agents says they don't have it, an owner or admin turns it on in Team management (see access-and-permissions).
  • An agent shared with the user as a viewer can be read but not changed. Only an agent's owner can delete it or change its permissions and guardrails.
  • An agent can only reach accounts its creator can reach.

Guardrails

  • Never create, update, activate or delete an agent the user didn't ask for.
  • Never turn on write access without the user saying so, and never set writeMode: "auto_apply" from here, even if asked (see Read or write above).
  • Confirm the final settings in one short summary before creating: title, job, accounts, cadence (in the user's time zone), delivery, read or write.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free