Prompts and resources

Beyond tools, the server ships two kinds of guidance: prompts and resources. Both encode the opinionated workflows behind the tools, so a client that has never seen the data still runs the analysis the way an expert would.

What prompts are

A prompt is a pre-written instruction message the server hands to your client when you invoke it. Invoking a prompt injects an expert checklist into the conversation — the tool sequence to run, the watchouts to apply, and the order that matters — and your client then executes it with the tools. Prompts take no arguments; they act on the session’s current account (or whatever accountId you pass).

What resources are

A resource is a markdown document the server serves at an aga:// URI. Your client can read one mid-task — the way it would read a file — to load the long version of a rule set without leaving the session. The three resources are the operational condensation of this site’s guides.

How to invoke them

Claude Code — server prompts appear in the / slash menu; pick one to run it. Resources can be read by their URI.
Cursor and Claude Cowork — prompts appear in the client’s UI for the connected server; invoke them from there.
Any agent — you can also just tell it: “use the content-deep-dive prompt on this account.” The client fetches the prompt message and follows it.

Fallback: everything the prompts and resources say is mirrored in these guides. If your client cannot invoke prompts or read resources, paste the relevant guide into the conversation — same content, same result.

The prompts

Five pairs plus wins-digest. In each pair, the general prompt crosses data from all the tools — it dispatches one subagent per data section (universe, citations and mentions, content, GSC) and synthesizes the briefs — while the gsc- variant runs the Google Search Console-only path.

Prompt	Use it when
`explore`	You are on an account you have never touched. Orients across the whole system before any performance question: publishing setup, the search universe, citations and mentions, and a first GSC overview
`gsc-explore`	You only need the GSC orientation: `gsc_get_context` first (reverse-proxy rewrites, property set, maturity cadence, placeholder flag), then `gsc_list_sites`, then `gsc_get_overview` read impressions-first
`performance-check`	You want the periodic “how is the account doing” read across the full loop: search universe, citations and mentions, published content, and GSC movement, crossed together
`gsc-performance-check`	You only need the GSC read. Picks the cadence from account maturity (week-over-week only in the first ~8 weeks), runs `gsc_compare_periods`, and separates the impression growth share (AI visibility) from the click share (human traffic)
`wins-digest`	You want “what did we win this week” on Google. Snapshot dates from `reporting_get_available_dates`, then `reporting_get_wins`, then planner drill-down. Google-only — never mixes in ChatGPT modes
`content-deep-dive`	You are mapping the account’s full system: search universe, published corpus, citations and mentions per term, and GSC performance per page — crossed together
`gsc-content-deep-dive`	You are running the GSC-only deep analysis. The complete runbook: site structure before cohorts, quota budgeting, monthly exports under the 25k cap, the anonymization block before any query-level claim
`page-diagnosis`	You want one page’s full footprint: what it earns in Google search and where it is cited or mentioned across the universe, ending in a keep / refresh / retire / replicate call
`gsc-page-diagnosis`	You only need the GSC page read. What queries the page earns — including demand beyond the query it was published for, its blast radius — then `gsc_get_page_performance` → `gsc_inspect_url` (with the canonical-URL caveats) → `gsc_list_sitemaps`
`trend-explanation`	A number moved — a citations/mentions rate or a GSC metric — and you need to know why: universe-size change, recent publishes, dilution vs decay, net vs gross churn
`gsc-trend-explanation`	A GSC number moved. `gsc_compare_periods` movers plus the AGA split, dilution vs decay, net vs gross churn, and the universe-size check before any rate read

The three resources

Resource	Read it when
`aga://guides/first-analysis`	Before running a deep analysis on a new account — the Phase 0–5 runbook as a checklist
`aga://guides/gsc-data-fidelity`	Before trusting any figure from the `gsc_*` tools — what the instrument does to the data (anonymization, lag, row caps, quota, remote files)
`aga://guides/analysis-guardrails`	Before modelling anything from search data — the analytical patterns that keep the analysis honest (de-trending, balanced panels, placebos, dilution vs decay)

Client-side skills

Prompts and resources live on the server. If your agent harness supports installable skills (a SKILL.md with a name and description), you can also install small wrappers that trigger the same workflows from your own side. A skill description should state only the triggering conditions — start it with “Use when…” — and the body should stay a short checklist that points at the guides rather than duplicating them. Two examples:

---
name: aga-gsc-weekly-check
description: Use when asked for a periodic account check-in, a weekly
  performance read, or a wins digest on an AI Growth Agent account.
---

- Invoke the `performance-check` prompt (or follow
  https://docs.aigrowthagent.co/guides/prompts-and-resources).
  Cadence comes from account maturity: week-over-week only in the first
  ~8 weeks of publishing, month- or quarter-over-quarter after.
- Invoke the `wins-digest` prompt for the week's new citations and brand
  mentions. Google-only — never mix in ChatGPT modes.
- Read impressions first: impression growth share is the AI-visibility
  read; the click share is human traffic. Never present a click move as
  an AI-visibility move.
- Read every rate change against universe size: a dip against a growing
  universe is expansion, not decline.

---
name: aga-gsc-deep-analysis
description: Use when asked to run a full content-performance analysis,
  an audit, or any modelling work over Google Search Console data on an
  AI Growth Agent account.
---

- Invoke the `gsc-content-deep-dive` prompt and follow
  https://docs.aigrowthagent.co/guides/first-analysis phase by phase:
  site structure before cohorts, budgeted acquisition, panel construction,
  ground truth — then model.
- Pull `[page, query]` early and read `meta.anonymization` before any
  query-level claim — named queries are a biased floor
  (https://docs.aigrowthagent.co/guides/gsc-data-fidelity).
- `curl` every export `downloadUrl` immediately — files are remote and
  the signature expires in ~1 hour.
- Before modelling, apply the guardrails:
  de-trend every age curve, use balanced panels, run placebos before
  believing any event study, and separate dilution from decay
  (https://docs.aigrowthagent.co/guides/analysis-guardrails).

Install either file under your harness’s skills directory and it triggers on the matching requests. Keep them thin — when the workflows evolve, the server prompts and these guides are the source of truth, not the skill body.

Get started

Tasks

Guides

Reference

What prompts are

What resources are

How to invoke them

The prompts

The three resources

Client-side skills

See also

​What prompts are

​What resources are

​How to invoke them

​The prompts

​The three resources

​Client-side skills

​See also

What prompts are

What resources are

How to invoke them

The prompts

The three resources

Client-side skills

See also