Blocks

Summarize

Condense input to the essentials at short, medium, or long length. Where the summary lands in the payload, digest patterns, and when a prompt fits better.

Condense the input to the essentials. Summarize is text-in, prose-out, the standard last step before a send email or send text.

What it does

Summarize is the AI Agent step with its Task set to summarize. You don't write a prompt. The step synthesizes one from your settings ("Summarize the input to the essentials", the length rule, your extra instructions, then the input) and runs it through the same agent pipeline a custom prompt uses. The model is the platform model unless you bind a Provider (see AI settings).

The step always runs in text mode. There is no output schema, so the model's prose is the result. That is deliberate: a summary is the end of a chain, not a source of fields.

Config

The summarize task in the Inspector Length, and optionally what to read and extra guidance.

  • Task: summarize. Switching it changes the fields below.
  • Length: short (one or two sentences), medium (one tight paragraph, the default when unset), long (a few paragraphs when the input warrants it).
  • Input: optional mustache template picking what to read, for example {{ payload.body.text }}. Blank reads the whole upstream payload, JSON-encoded.
  • Extra instructions: optional guidance folded into the prompt, "bullet points, lead with any deadline".
  • Provider and Model: empty runs on the platform model. Bind an AI connection to use your own key; the Model field is honored only with a connection.
  • Temperature (Advanced): 0 to 1. Blank leaves the provider default; unlike classify, summarize does not pin it.
  • On failure (Advanced): fail (default) stops the run, retry re-runs the step (0 to 5 retries with a backoff), and continue routes an {ok: false, error} payload out the error port so the workflow survives.

Output

Text mode: the summary lands at payload.text:

{
  "ok": true,
  "text": "Three customers reported checkout failures…",
  "raw": "Three customers reported checkout failures…",
  "usage": { "input_tokens": 812, "output_tokens": 96, "total_tokens": 908 },
  "error": null
}

Downstream references {{ payload.text }}. raw is the same string, usage carries token counts, and error is null on success. There are no named fields to pluck, and the build-time linter treats this step as prose: a later step that references payload.data.something from a summarize step is flagged.

Patterns

  • Digest pipeline: schedule or watch gathers items, reshape aggregates them, summarize condenses the lot, send email delivers it.
  • Long-thing-to-short-thing: a webhook delivers a full document; summarize before posting to chat.
  • Per-item summaries: inside a loop, set Input to {{ payload.item.body }} so each iteration summarizes one record.

Gotchas

  • Summarize is terminal-shaped: downstream steps can only reference the one payload.text string. If a later step needs fields (a title AND a body), use a custom prompt with a schema instead.
  • Point Input at the actual content when the payload carries metadata too. A blank Input serializes the whole upstream map, headers and all, and the model spends attention on plumbing.
  • Length is a steer, not a cap. A short summary of a long input is still a sentence or two, but the model decides where the line falls. Add an instruction if you need a hard shape such as "three bullets".
  • The step fails the run on an API error by default. Set On failure to continue and wire the error port when a digest should still send with a fallback line.

Related

‹ Classify Set fields ›