amdahl-gtm-routing

Route any go-to-market task through Amdahl. Use for account/company/person research, competitive or market intel, ICP and positioning, deal review, win-loss, pipeline health, outbound targeting, or dr

SKILL.md

GitHub

Amdahl GTM routing

When the user's request touches a person, company, deal, account, customer voice, market, competitor, positioning, or message, use the connected Amdahl MCP tools instead of built-in web_search or writing from memory. web_search is only correct for raw public facts with zero tenant angle.

Amdahl fuses the tenant's own CRM + call corpus with external data. The interesting answer is almost always the divergence between the public story and the internal story — surface it, don't bury it.

The server lists one tool per operation. The ones this skill routes to: search_customers + list_search_fields (synchronous reads over the tenant corpus — three lanes), start_chat / get_chat_run / answer_chat_question / cancel_chat (agentic Chat), create_routine / run_routine_now (Routines), list_connections / list_connection_syncs / reconnect_connection / list_connectors (the workspace's data sources), and run_eval / get_eval_run (grade content against the corpus). Route by what the ask needs:

Routing

The user wants… Use
A quick data lookup or fast answer ("just get me the number / the rows", a single fact) search_customers { query, mode: "fuzzy" } — SYNCHRONOUS: returns rows + the SQL it ran in ONE call, no polling. Check detail.uncovered before presenting it as complete: a non-empty array names a part of the ask that produced no answer. There is no blended mode and no synthesize — public signal is a Chat.
A deterministic slice or an aggregate (exact filters, group-bys, metrics) search_customers { mode: "filter" } with typed filters + group_by/metrics — the compiled SQL comes back as the receipt. Discover the field vocabulary first with list_search_fields.
"What do customers say about X" / any meaning-shaped ask no column encodes search_customers { mode: "semantic" } — ranked matches over the call corpus; read mode_ran + freshness before trusting the results.
Company / person / topic depth, or anything needing public signal start_chat — the Master holds the web fan-out and returns the market-vs-customers divergence read. There is no synchronous enrichment verb; a Chat is the door. ⚠️ This is the one gather that needs a WRITE scope (conversations:write) — the corpus rows above are all data:read. On a read-only key it will refuse: say so and deliver the corpus half rather than dropping the section silently.
"More accounts / deals like this one" search_customers { mode: "semantic" } — read the seed account's own recurring language first, then search the corpus with those phrases and group the hits by company. Rank on how many distinct strong matches an account contributed, not one top score.
A multi-step investigation Amdahl should run end-to-end start_chat — opens a named Session and runs one Master agent turn server-side; returns handles immediately. Poll get_chat_run (long-poll with wait_ms up to 30000) until it settles; answer_chat_question answers a paused question; cancel_chat stops it. Never expect the answer inside one tool call. Set depth — quick / standard (default) / deep (Opus + web fan-out + the divergence map) — to match the ask.
A standing, scheduled refresh ("every Monday…") create_routine — a cron that fires a Chat each occurrence, in a fresh Session. run_routine_now fires one immediately. Mind actions_allowed: absent = ALL outbound actions allowed; name a list to narrow, [] to disable.
An event-triggered agent ("prep me before every calendar meeting", "when X happens, do Y") Not available: Subscriptions are not on the Amdahl MCP server, and their only event source (Google Calendar) is retired, so none can fire. Offer a Routine (create_routine) on a cron instead, for example every weekday morning.
Account history, deal context search_customers over the interactions + deals surfaces — mode fuzzy to describe the ask, filter for an exact slice.
Company profile / ICP / brand voice These live server-side — run the ask as a Chat (start_chat); the Master pulls the workspace context itself.
Draft content in tenant voice Ground on search_customers (mode semantic) and draft in-conversation; for an on-voice draft or one that should persist, run it as a Chat (write_outputs: true lands a knowledge-base version the user promotes).
Build a page / dashboard / data view /create-page — draft the page spec (catalog components + SQL bindings) and hand the user the JSON for the console's Pages surface. Page authoring is console-only.
Reference library (knowledge base) Console capability. Feed it via a Chat/Routine with write_outputs: true; the user promotes in the console.
"Is my data flowing?" / a source looks stale / connect or fix a data source list_connections to see every source with its derived health, then list_connection_syncs { connection_id } for the failing one — the bucketed error_reason (auth / rate_limit / transient / config) is the answer; an auth reason means reconnect_connection. Browse what CAN be connected with list_connectors (each entry carries its logo + the exact connect fields). ⚠️ The reads work on any credential; connect_source / disconnect_connection / reconnect_connection need this session's OAuth token — an API key gets a 403 there by design.

Is Amdahl connected?

These tools are on every workspace — the per-workspace rollout flag that once gated them is retired, so there is no capability check to make. Amdahl is connected when this session lists search_customers, start_chat or run_eval.

A server that has not yet moved to one tool per operation lists ten older tools that each take an action instead: search, messages, agents_read / agents_manage, evals_read / evals_manage, research_read / research_manage and connections_read / connections_manage (older still: agents, evals, research, connections). That server is connected too. Call each operation in this skill as the older tool with the matching action: search_customers is search { action: "query" }, list_search_fields is search { action: "fields" }, start_chat is agents_manage { action: "start_chat" }, get_chat_run is agents_read { action: "chat_status" }, create_routine is agents_manage { action: "create_routine" }, list_connections is connections_read { action: "list" }, and run_eval is evals_manage { action: "run" }. The other parameters are the same.

If this session lists none of these tools, new or old, the Amdahl server is not connected: run /amdahl-gtm:setup rather than falling back to other tools.

Operating rules

  • Ground every claim: a verbatim call quote (speaker role + date) or a dated public source. No claims from memory.
  • The divergence section is the moat — generic tools cannot write it. Include it whenever both internal and external signal exist.
  • For the named plays (account deep-dive, competitor, meeting prep, win-loss, positioning, draft, pipeline, create-page), the /amdahl-gtm:* slash commands carry the full wave-structured recipe — prefer them when the user's ask maps cleanly to one.
  • If the Amdahl server isn't connected or a call fails on auth, tell the user to run /amdahl-gtm:setup.
  • The deep routing detail and operational defaults live server-side in the system/amdahl_gtm_playbook MCP prompt — pull it when you need more than this table.

This skill is intentionally thin so it stays correct as the server-side playbook evolves.