Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.51ultron.com/llms.txt

Use this file to discover all available pages before exploring further.

Every Ultron session starts with context already loaded. By the time you type your first message, the model already knows your business, your ICP, your voice, your connected tools, and the most relevant memories from past sessions.

​
Injection order

Session start — system prompt assembly
├── 1. Static instructions
│      Ultron behavior rules, tool use guidelines,
│      quality gates (VIRALS scoring, cold email rules),
│      anti-fabrication rules, response style
├── 2. User profile
│      Company name, ICP definition, voice tone,
│      competitors, platforms, voice samples,
│      banned words
├── 3. Connected integrations
│      List of active services: Gmail, Stripe,
│      HubSpot, Notion, Slack, etc.
├── 4. Relevant memories (max 5)
│      Selected by Sonnet-powered recall query
│      against current session objective
└── 5. Current date
           Injected as system reminder per turn

​
Why order matters

The ordering is deliberate, not arbitrary. Static content at the top is cache-friendly — it doesn’t change between sessions, so the model can reuse previous computation. Dynamic content at the bottom always gets fresh context. This means:
  • Faster session start times (cache hits on static sections)
  • Lower token cost for repeated sessions
  • Fresh profile and integration data every time

​
Profile completeness

The quality of Ultron’s output is directly proportional to how complete your profile is.
Profile fieldImpact on output
ICP definitionSpecter targets the right leads. Striker personalizes correctly.
Voice tonePulse writes in your style. Striker emails sound human.
Voice samplesThe single highest-leverage input. Actual examples = accurate mimicry.
CompetitorsCortex frames research correctly. Striker handles objections accurately.
Banned wordsEvery output automatically avoids your banned patterns.
Add 2–3 voice samples — real emails or LinkedIn posts you’ve written. This one change improves output quality more than any other setting.

​
Anti-fabrication rules

Part of the static instructions injected at session start. If a web search returns no data on a specific pricing figure or revenue number, Ultron uses qualitative language (“strong growth signals”, “market-leading position”) rather than inventing a number. The instructions explicitly prohibit fabricating stats, pricing, or revenue figures.