When Neural Summary launched, it had seven analysis types: summary, communication styles, action items, emotional intelligence, influence and persuasion, personal development, and custom. They were generic. The same prompt structure, with minor variations, applied to every conversation type.
Eight months later, we have 30 lens templates organized into five intent-based categories. Each one is designed by positioning the AI as a specific domain expert. The output quality is incomparable to where we started.
This post covers how we think about template design, the value ladder that organizes them, and what we learned about turning generic AI output into professional-grade deliverables.
The value ladder
Templates are organized into five categories that progress from passive to active. We call this the value ladder because each step produces more immediately actionable output.
Capture — Structure the raw conversation. Meeting minutes, 1:1 notes, coaching session notes. The output tells you what happened. This is where most meeting tools stop.
Analyze — Extract patterns and intelligence. Communication analysis, competitive intelligence, deal qualification, retrospective, process map. The output tells you what it means.
Communicate — Turn insights into outreach. Follow-up emails, sales emails, client proposals, LinkedIn posts, newsletters, internal updates. The output is something you send to someone.
Deliver — Generate work-ready deliverables. Blog posts, product requirement documents, strategy briefs, agile backlogs, technical design documents, case studies, presentation outlines. The output is a finished artifact your team can act on.
Act — Drive decisions and next steps. Action items, decision documents, executive summaries. The output translates conversation into motion.
The progression is deliberate. Capture requires the least interpretation. Act requires the most. Each step up the ladder produces output that saves more time and requires more sophisticated prompt design.
Anatomy of a lens template
Every template has five components:
1. Domain expert positioning. The opening instruction that sets the AI's perspective. Not "summarize this transcript" but something shaped like "you are a seasoned practitioner producing a document a client would pay for." The role carries the quality bar: the more specific the expertise, the more professional the output.
2. Output schema. A JSON schema defining every field, its type, and whether it is required. A backlog-style lens, for example, defines nested structure: epics that contain stories that contain acceptance criteria, plus the supporting context around them. The schema is what makes the output renderable, searchable, and translatable.
3. Quality examples. Good and bad examples for the fields that matter most. For user stories: a good example has a specific persona, a concrete action, and a measurable outcome. A bad example has "as a user, I want to do things."
4. Constraints. Word limits, format requirements, and structural rules: list items capped at a fixed word count, action items forced to start with an imperative verb, headings held to a few words. These constraints are what prevent the LLM from producing verbose, meandering output.
5. Edge case instructions. What to do when the transcript does not contain enough information. If the conversation never touched budget, the output should say so rather than invent a number. This prevents hallucination in sparse transcripts.
The evolution: V1 to V2
The difference between our early templates and our current ones is instructive. The prompts below are illustrative sketches, not our production prompts; the real ones run far longer, and they are the product.
V1 Executive Summary (October 2025):
Prompt: "Summarize the key points of this meeting.
Include main topics, decisions, and action items."
Output: Markdown blob with bullet points.
V2 Executive Summary (March 2026):
Prompt: "You are an experienced management consultant
preparing a decision-ready brief for a steering
committee. Lead with the one recommendation the
group should act on. Tie every finding to evidence
from the conversation and spell out its implication.
Record each decision with an owner and a rationale,
and each risk with a severity and a mitigation."
Output: Typed JSON with recommendation, findings, decisions,
risks, and action items, every field validated
against a schema.
The V1 output read like a student's meeting notes. The V2 output reads like something a management consultant would produce for a client.
The difference is not the model. We tested V1 prompts on newer models and V2 prompts on older models. The prompt is the dominant factor.
Case study: CRM Notes
The CRM notes template had the most dramatic transformation.
V1: A narrative summary with some bullet points about what was discussed.
V2: Sales intelligence in a structured format. A one-line deal verdict. Qualification against a standard sales framework, with a status per dimension instead of prose. A stakeholder map covering each person's role and influence. Pain points ranked by business impact. Buying signals and objections tied to what the prospect actually said, each objection paired with a suggested response. Competitors mentioned, with the context they came up in. Next steps grouped by timeline, with owners.
A sales professional using V1 got meeting notes they could paste into a CRM. A sales professional using V2 gets a complete call intelligence brief that structures their follow-up strategy.
Case study: Agile Backlog
The agile backlog template shows how domain expertise in the prompt produces radically different output.
The V1 prompt generated a flat list of user stories. They were grammatically correct but generic: "As a user, I want to create an account so that I can access the platform."
The V2 prompt positions the AI as a senior product practitioner who has shipped real backlogs to real development teams. It requires personas extracted from the conversation instead of generic "user" placeholders, stories with testable acceptance criteria, relative sizing, epic grouping with priorities, and explicit assumptions and out-of-scope items.
The prompt includes specific examples of good and bad acceptance criteria, in the spirit of: good is "given a cart with three items, when the shopper removes one, then the order total updates without a page reload"; bad is "the system should update correctly."
The output is a backlog that a development team can import and start working from, not a list that requires a product manager to rewrite every story.
The template design process
New templates follow a consistent process:
1. Interview the expert. What would a senior professional in this role actually produce? We study real examples: actual consulting briefs, actual sales intelligence reports, actual coaching session documentation. The template should produce output that matches professional standards, not a summarized version of them.
2. Define the schema. What fields does the output need? What types are they? Which are required versus optional? The schema is the contract between the AI and the rendering layer.
3. Write the prompt. Position the expert, define the schema, provide examples, set constraints. The first draft is never good enough.
4. Test against sample transcripts. We maintain a set of representative conversations across different types (sales calls, coaching sessions, team meetings, strategy discussions). Every template is tested against this set.
5. Review output quality. Would the intended user be proud to send this to a client? If not, iterate on the prompt. Most templates go through 3-5 iterations before reaching production quality.
6. Add backward compatibility. If we are redesigning an existing template, the new renderer must handle both old-format and new-format data. Users have existing lenses generated with the old schema. They must continue to render.
What we learned
The prompt is the product. Model selection matters, but prompt design matters more. A well-structured prompt with domain expertise, examples, and constraints produces consistently better output than a generic prompt on a more powerful model.
Backward compatibility is non-negotiable. Users have existing documents. When we redesign a template's schema, old data must still render. Every new field is optional. Every renderer handles the absence of new fields gracefully.
Domain expertise beats general intelligence. Every template that moved from a general-assistant perspective to a specialist one improved dramatically: a coaching lens grounded in how people actually develop, a competitive lens written from the desk of a seasoned analyst rather than a note-taker. Specificity in the prompt produces specificity in the output.
The value ladder is a product strategy. Most meeting tools compete in the Capture category. We invest our template development time in Deliver and Act, where the time savings per lens are highest and the competition is thinnest.
Thirty templates. Five categories. Each one designed to produce output that a professional would actually use, not just read.
That is the standard. And we are not done.
