Project 02 · Domain AI
Domain-Specific AI Expert
A Gemini Gem constrained to one vehicle, not the average Gladiator.
Applied — personal vehicle domain
Mission
Replace generic vehicle answers with an assistant that actually knows the trim, drivetrain, tow package, and owner context.
Problem
Generic AI answers about a vehicle ignore the specifics that actually matter — trim, drivetrain, tow package, tire and gear setup, prior modifications, service history.
Constraints
- No direct write access — the assistant recommends, it never executes.
- Purchases, modifications, and anything safety-relevant require owner sign-off.
- Forum consensus must not be presented as verified specification.
My role
Designed the system prompt and domain rules, curated the owner-context and vehicle-configuration inputs, and defined which categories of output required my own sign-off versus what could be treated as plain reference information.
Architecture
Separated owner context, vehicle configuration, and maintenance/technical reference into distinct inputs. Built a persistence pattern so context did not need to be re-explained every session. Instructed the assistant to flag judgment calls differently from settled facts.
Owner context
Reference answer
Vehicle configuration
Domain rules
Gemini Gem
Configured assistant
Technical reference
Owner decision
Purchases, mods, safety
- Owner context
- data
- Vehicle configuration
- data
- Technical reference
- data
- Domain rules
- policy
- Gemini Gem
- agent — Configured assistant
- Reference answer
- output
- Owner decision
- approval — Purchases, mods, safety
AI architecture
- Model
- Gemini Gem
- Agent
- None — no tool selection, no investigation state, no actions
- Context
- Owner, vehicle configuration, and reference kept distinct
- Human interaction
- Sign-off on purchases, modifications, safety-relevant output
Security architecture
- No write access
- Recommend-only capability
- High-impact routing
- Human decision for safety-relevant change
- Source verification
- Do not treat unverified consensus as specification
Build process
- System prompt and domain rules designed around one specific vehicle.
- Owner-context and configuration curated as durable inputs.
- Output categories split into reference versus sign-off required.
Validation
Recommendation cross-check
Manufacturer service documentation and parts compatibility references
No recommendation acted on without that check
| Check | Method | Result |
|---|---|---|
| Recommendation cross-check | Manufacturer service documentation and parts compatibility references | No recommendation acted on without that check |
Challenges
A generic chatbot will happily answer as if every Gladiator is the same vehicle. The failure is not intelligence — it is missing constraints.
Investigation
Compared generic answers against manufacturer documentation. The gap was context engineering and constraint definition, not model choice.
Resolution
Persist structured context, separate facts from judgment, and keep execution out of the assistant’s reach.
Lessons
Name the class honestly
This is a configured assistant, not an agent. It does not select tools, hold investigation state, or take actions. Being precise about that distinction is what made the later agent designs possible.
Cybersecurity equivalent
- Recommend-only assistantmaps toNo standing change privilege
- Safety-relevant sign-offmaps toPrivileged change authorization
- Source-verification instructionmaps toProvenance control
AI principles applied
LLMs as bounded components
Models draft, classify, and summarize. They do not own authority, publication, or remediation.
Where learned: First as specialized GPTs on a service-desk publishing workflow, then as Gemini inside TinyClaw and a domain Gem.
Prompt and role configuration
Persistent role, rules, and output contracts — not a one-shot instruction.
Where learned: Applied as a specialized Gemini Gem for a specific vehicle-maintenance domain.
Context engineering
Separate owner context, configuration, and reference instead of blending them into one prompt.
Where learned: The Jeep assistant made this the whole product.
Trusted retrieval / source timing
Normalize and timestamp sources before the model sees them. Forum consensus is not a specification.
Where learned: TinyClaw freshness path; Jeep source-verification instruction.
Concepts exercised in this project
- Configured assistants
- Context engineering
- Constraint definition
- Source verification
Security principles applied
Controls exercised in this project
- Human approval
- Least privilege
- Provenance
Technologies
- Gemini Gem
- System prompting
- Domain rules
Skills demonstrated
- Domain-Specific AI Design
- Context Engineering
- System Prompting
- Constraint Definition
- Source Verification
- Specialized Assistant Design