Best practices for building high-accuracy custom summary templates?

General
1

Hey everyone — I’m trying to push Pocket’s custom templates pretty hard and wanted to ask for some guidance on how the template system actually interprets prompts.

I’m using Pocket as a long-term memory layer / Chief of Staff for real conversations: business calls, strategy sessions, social conversations, family discussions, and idea capture. My goal is not just a clean summary — I’m trying to preserve accurate speaker attribution, decisions, action items, relationship updates, useful context, and “what to watch next” without the output becoming too long or essay-like.

I’m currently testing a condensed custom template with sections like:

  • Main Summary
  • Key Takeaways
  • Relationships & Thoughts
  • Action Items & Open Loops
  • Predictions & Signals

The main things I’m trying to understand are:

  1. When using a custom template, does Pocket treat every section as mandatory, or can the model naturally shorten/skip sections depending on the conversation?
  2. Does Auto Detect use a different internal prompt or structure than custom templates?
  3. Are custom templates interpreted section-by-section, or as one full prompt?
  4. Is there a best practice for getting more accurate speaker/entity attribution, especially when a transcript includes active speakers, background voices, and people being discussed?
  5. Can templates reliably use instructions like “label plans as confirmed, tentative, discussed, or unclear”?
  6. Are XML-style blocks like <pocket:timeline> officially supported in custom templates, or are those only generated by built-in themes?
  7. Is there an ideal character length for each section prompt before quality starts dropping?
  8. Does Pocket use imported profile/memory context during template generation, and if so, how heavily does it weigh that compared to the transcript?

My ideal output is something close to Auto Detect’s adaptive structure, but personalized: concise, accurate, skimmable, and smart enough to capture relationship dynamics, action items, and future signals without over-explaining.

One feature request: I’d love a pre-summary “cast list” or speaker/entity review step where I can confirm:

  • active speakers
  • background voices
  • people discussed
  • uncertain names/entities

That would make long-term memory and relationship tracking much more accurate.

Would love any guidance from the Pocket team or power users on how to structure templates for the best results.

4 Likes