Introducing a Content Density Mode System

Status: Draft Date: 2026-03-10 Last Updated: 2026-03-10 Author: AI Labs Team Related: Introducing-a-Table-Generator-Agent.md, Post-Generation-Quality-Agents.md, Format-Memo-According-to-Template-Input.md

Executive Summary

Different firms have fundamentally different philosophies about memo density and repetition. Some assume readers skim — they want key data points repeated across sections so a reader landing on any section gets context without reading the whole document. Others read front-to-back and find repetition tedious and unprofessional.

This specification introduces a content_mode setting ("redundant" or "concise") that propagates across multiple agents, altering how they handle information repetition, table/prose relationships, and cross-section references.

The Two Philosophies

Redundant Mode (Default)

Assumption: Readers skim. Partners flip to the section they care about.

Behavior:

• Key metrics (funding, ARR, team size) appear in multiple sections where relevant
• Tables supplement prose — the same data appears in both formats
• Executive Summary contains substantial detail, not just references
• Cross-section references are rare; each section is semi-self-contained
• Longer overall memo length is acceptable

Who wants this: Firms with large partnerships where different partners read different sections. Firms producing memos for LPs who will skim. Firms where memos serve as reference documents.

Concise Mode

Assumption: Readers read front-to-back. Repetition wastes their time.

Behavior:

• Each data point appears once, in its most relevant section
• Tables replace inline enumerations (prose is trimmed to narrative analysis only)
• Cross-section anchor links connect related information: “The team (see Organization) brings…”
• Executive Summary is tight — references body sections for detail
• Shorter overall memo length

Who wants this: Firms with small teams who read memos thoroughly. Firms where memos are decision documents, not reference documents. Users who reported current output is “too long” or “too repetitive.”

Configuration

In Outline YAML (Preferred)

# templates/outlines/direct-investment.yaml
metadata:
  name: "Direct Investment Memo"
  content_mode: "redundant"  # or "concise"

In Company JSON (Per-Deal Override)

{
  "type": "direct",
  "content_mode": "concise"
}

Resolution Order

1. Company JSON content_mode (per-deal override) → highest priority
2. Outline YAML content_mode (firm default)
3. System default: "redundant"

Agent Behavior by Mode

Writer Agent

Implementation: The writer’s system prompt includes a content mode directive:

CONTENT MODE: {content_mode}

If REDUNDANT:
- Each section should be readable on its own. A partner flipping to "Team"
  should get enough context without reading prior sections.
- Repeat key metrics where they add context (funding in both Exec Summary
  and Funding & Terms, team highlights in both Business Overview and Team).
- Err on the side of completeness per section.

If CONCISE:
- State each data point once, in its most relevant section.
- Use anchor links to connect related information across sections.
- Keep prose analytical, not enumerative. Let tables handle data display.
- Target the lower end of word count ranges.

Table Generator Agent

Example — Concise mode transformation:

Before (prose):

Key competitors include Calm, which has raised $218M and serves 4M subscribers,
Headspace with $215M raised and 2.5M subscribers, and Noom at $540M raised
with 1M subscribers. The competitive landscape is intensifying as consumer
wellness spending grows.

After (concise mode with table):

The competitive landscape is intensifying as consumer wellness spending grows.

| Competitor | Funding | Subscribers |
|------------|---------|-------------|
| Calm       | $218M   | 4M          |
| Headspace  | $215M   | 2.5M        |
| Noom       | $540M   | 1M          |

After (redundant mode with table):

Key competitors include Calm ($218M raised, 4M subscribers),
Headspace ($215M, 2.5M subscribers), and Noom ($540M, 1M subscribers).
The competitive landscape is intensifying as consumer wellness spending grows.

| Competitor | Funding | Subscribers |
|------------|---------|-------------|
| Calm       | $218M   | 4M          |
| Headspace  | $215M   | 2.5M        |
| Noom       | $540M   | 1M          |

Redundancy Reducer Agent

Revise Summary Sections Agent

State Propagation

The content mode is set once and flows through the entire pipeline via state:

# In MemoState (src/state.py)
class MemoState(TypedDict):
    # ... existing fields ...
    content_mode: str  # "redundant" or "concise"

Set during initialization in src/main.py:

content_mode = company_data.get("content_mode") or outline.get("metadata", {}).get("content_mode", "redundant")
state["content_mode"] = content_mode

Each agent reads state["content_mode"] and adjusts behavior accordingly.

Implementation Priority

This is a lower priority than the competitive landscape system and table generator. The content mode system is a refinement that improves output quality for different firm preferences, but the core agents work without it (defaulting to redundant mode, which is current behavior).

Suggested order:

1. Add content_mode field to state and config loading (small change)
2. Wire into writer agent prompt (moderate change)
3. Wire into table generator (moderate change)
4. Wire into redundancy reducer (when that agent is implemented)
5. Wire into revise summary sections (small change)

Open Questions

1. Should there be a “balanced” middle option? Or is the binary toggle sufficient? (Recommendation: start binary, add a middle option later only if needed.)

2. Per-section content mode? Some firms might want the Executive Summary to be redundant (self-contained) but the body sections to be concise. This adds complexity — defer unless specifically requested.

3. How does this interact with the outline’s word count ranges? Currently each section has min_words and max_words in the outline. Concise mode would target min_words, redundant targets max_words. Should we add explicit concise_target_words and redundant_target_words?