help-docs

name	help-docs
description	Live retrieval of Make.com product documentation from help.make.com, with a local cache. Answers conceptual questions — routers, iterators, error-handler directives, operations and billing, AI Agent semantics, plan limits — that the MCP module docs do not cover. Always cites the source URL.
allowed-tools	Read, Write, Bash, Glob, Grep, WebFetch, mcp__firecrawl__firecrawl_search, mcp__firecrawl__firecrawl_scrape

Skill: help-docs

Fetches Make.com product documentation from help.make.com on demand and caches it locally, so conceptual answers come from the current docs instead of from memory.

Scope boundary — read this first

Two doc sources exist and they do not overlap. Using the wrong one is the failure mode this skill prevents.

Question	Source	Why
Exact module field names, parameter types, required vs optional, enum values	MCP app-module_get, app_documentation_get	Authoritative — reflects the live connector schema
Concepts, semantics, limits, billing, behaviour: routers, iterators/aggregators, error-handler directives, operations counting, data stores vs data structures, AI Agents, plan limits, scheduling	help.make.com (this skill)	Not exposed by the MCP module API

MCP wins on any conflict about a field spec. Prose on help.make.com is often written against an older connector version. Never override a live module schema with help text.

Retrieval sequence

1. Cache first

Look in .make/research/help/ for a matching topic:

ls .make/research/help/ 2>/dev/null | grep -i "<topic-keyword>"

Each cached file carries frontmatter:

---
topic: error-handling-overview
source: https://help.make.com/overview-of-error-handling
page_updated: 21 Aug 2026
fetched: 2026-08-24T14:02:00Z
---

If fetched is less than 14 days old, use it and make no network call. Say you answered from cache and give the source URL.

2. Cache miss or stale → fetch

Use includeDomains, not a site: operator — it filters reliably and the operator does not:

mcp__firecrawl__firecrawl_search
  query: "<topic in plain words>"
  includeDomains: ["help.make.com"]
  limit: 5

Pick canonical URLs from the results. help.make.com search results include query-parameter-polluted duplicates of the same article (...?OLSCode=...&kai=... — hundreds of junk params); strip the query string and keep the bare path. Some article URLs are opaque ids (/64Ex54PN88vgyeJrsX10g) rather than slugs — that is normal, keep them as given.

Then scrape the 2–3 most relevant hits:

mcp__firecrawl__firecrawl_scrape
  url: <canonical url>
  formats: ["markdown"]
  onlyMainContent: true

onlyMainContent: true is not enough on this site. The docs run on Archbee and every page still returns the full left-hand nav (~80 links), a trailing table of contents, and "Docs powered by Archbee" boilerplate. Before caching, keep only from the page's first # heading down to the Updated <date> line, and drop the rest. Caching the nav wastes most of the file on links you already know.

Useful fields on the response: metadata.title, metadata.sourceURL, and the Updated <date> line in the body — the last is Make's own revision date and belongs in the cache frontmatter as page_updated.

3. Write the cache

Slugify the topic (kebab-case, no dates in the name — the frontmatter carries the date):

.make/research/help/{topic-slug}.md

Write the frontmatter above, then the scraped markdown. Overwrite a stale file in place; do not accumulate dated copies.

4. Answer

Return the relevant excerpt with the source URL. The URL is what makes the answer checkable — never drop it, never paraphrase a URL you did not fetch.

When to call this skill

• Any "how does X work in Make" / "what happens if" / "is X allowed on my plan" question
• Before designing a flow that leans on router, iterator, aggregator, or error-handler semantics — getting these wrong is the most common source of silent runtime bugs
• Before estimating operations or cost (billing rules change)
• When ai-agent-designer needs current AI Agent / tool-calling behaviour
• When failure-diagnostician finds no taxonomy match — search help.make.com before declaring a new failure pattern

Degradation — never fabricate

1. No mcp__firecrawl__* tools in the session → WebFetch on https://help.make.com/ (or a specific known article URL).
2. Neither available → say plainly: "I can't reach help.make.com in this session." Answer from the MCP module docs and label what is unverified.

Never invent a help.make.com URL, and never present remembered doc text as fetched. A cited URL that 404s is worse than no citation.

The docs are not infallible

Two observed failure modes, both of which mean verify against behaviour before betting a blueprint on a sentence:

• Pages occasionally carry unpublished editorial or draft text left in the body.
• Prose can lag the connectors. On any conflict about a field spec, the live MCP module schema wins — see the boundary table above.

Quote the docs for semantics; confirm with validate_blueprint_schema or a test run before relying on an edge case.

Output format

**From help.make.com** — {article title}
{source URL} · fetched {date} {· from cache}

{the answer, in plain language}

{caveat, only if the docs contradict what the MCP module schema says}