Use cases

Most useful flows are one of four shapes. Each shape below ships as a template (File → Start from template) and is published as a public flow you can open, inspect and fork without an account.

Read this page to pick a starting point; read Core concepts for the mechanics behind it.

Shape	Template	Public flow
One research pass → several deliverables	Industry intelligence	Open
Score something, then branch on the score	Lead qualification	Open
Upload a file → structured artifacts	Meeting → action plan	Open
Run on a timer and report, including “nothing”	Weekly monitor	Open

Fan-out — research once, produce several formats

Template: Industry intelligence report · Flow: sobm.it/rTkQAN9yst

One Deep Research node runs several search rounds and a synthesis pass. Its single output feeds three consumers in parallel: an Agent for the executive summary, a Table Generation node for the competitor comparison, and a Presentation node for the deck.

Why fan out instead of chaining: the three deliverables read the same bytes, so they cannot contradict each other. Producing them in three separate runs lets the wording drift.

What to watch:

• Narrow the input question. A broad topic produces confident generality — the model is not the limiting factor, the question is.
• Spot-check citations in the report. The research node returns sources; a link that merely sits next to a claim is the failure mode to look for.

The same shape also fits: one transcript → summary + action items, one product spec → landing page + FAQ + changelog entry.

Routing — force a number, then branch on it

Template: Lead research & qualification · Flow: sobm.it/knGv1Jzrs5

Research a company, then let an Agent in json-schema mode return a fixed structure — company, score 0–10, qualified, one sentence of reasoning, a few key facts. A Router branches on payload.score with gte 7: qualified leads get a generated HTML Page proposal, everything else ends at an Output.

Why the schema matters more than the model: a step allowed to answer in prose gives you “a strong candidate, though with some reservations” — which you then have to read and convert into a decision yourself. Asked for a number, the model commits to a number, and the Router can act on it.

What to watch:

• A Router takes exactly one branch and never fans out. Always give it an unconditional default edge, or a non-matching payload fails the run.
• Treat the score as a sort order, not a verdict. Read the reasoning sentence before acting on the number.

The same shape also fits: support-ticket triage, content moderation, invoice approval thresholds.

Upload → artifacts — turn a file into structure

Template: Meeting → action plan · Flow: sobm.it/5Kj2JFaLoL

An Upload File node carries the recording into Transcribe Audio. The transcript then fans out into an Agent summary and a Table Generation node that extracts action items with owner, due date and priority.

The instruction that makes it useful: tell the table node to leave the owner cell empty when nobody was named, instead of inferring one. An empty cell is a visible, arguable gap; a guessed name is a silent error.

What to watch:

• Audio must reach Transcribe Audio through its audio handle.
• A downloadable artifact (.xls, .pptx, .html) is produced per run and kept in the node’s artifact history.

The same shape also fits: contract → obligations table, invoice PDF → line items, CSV export → cleaned sheet.

Schedule — run on a timer and report, including “nothing”

Template: Weekly monitor → chat message · Flow: sobm.it/8LenVNVyyZ

A Web Search node checks the sources, an Agent decides in JSON whether anything actually changed, and a Router picks the message. Both branches end in an HTTP Request to a chat webhook. Set the cadence under Schedule runs — see Schedules.

Why both branches post: a monitor that only speaks up when it finds something is indistinguishable from one that has quietly stopped looking. If a source disappears behind a login, the messages keep arriving on time — they just stop containing it. This template therefore also asks the agent for a missing list of subjects it found no fresh data for.

What to watch:

• HTTP URLs are not {…}-templated. Put dynamic values in the JSON body with { "$from": "payload.headline" }.
• Scheduled runs are billed to the flow owner and appear in the normal run history.

The same shape also fits: uptime digests, regulatory-change watch, weekly KPI summaries.

Two shapes that disappoint

Worth knowing before you build them:

• A daily digest of everything. It runs reliably and gets read twice. A digest has to answer a question you actually have, otherwise it is a well-formatted notification you learn to ignore.
• A drafting assistant for decisions you have not made. Generating an email is fast; deciding whether to send it and what position to take is the slow part, and no node does that for you.

The rule both point at: automation pays where you were mechanically moving data between steps. Where the slow part was judgement, it adds a review step instead of removing work.

Related

• Core concepts — nodes, edges, handles, the run lifecycle.
• Quick start — build the smallest possible flow first.
• Router — conditions, operators and the default edge.
• Schedules — cadence, timezone, fixed input.