FG-0106 Self-Contained Learning
Description
Design collaboration environments where all necessary knowledge exists within the ecosystem. Agents should not need to search external sources for project context, conventions, or standards. Documentation, examples, and reference materials should be accessible within the working environment. This grounds the agent in authoritative project context, which reduces hallucination, ensures consistency, and creates a single source of truth.
Implementation
Maintain project documentation within the repository or working environment. Include architecture decisions, coding standards, and examples alongside the code. Create reference files that agents can read at session start. Avoid relying on the model's parametric (training-data) knowledge for project-specific facts; supply them as retrieved context instead.
Success Indicators
- New agents can onboard by reading project documentation alone
- Answers to 'how do we do X here' exist in the ecosystem
- Agent output is consistent with documented project standards
- Reduced hallucination about project-specific conventions
Failure Modes
- Documentation sprawl: too much documentation is as bad as too little
- Stale documentation that contradicts current practice
- Over-reliance on documentation instead of clear code and architecture
- Closed ecosystem bias: making the environment so self-contained that legitimate external knowledge and evolving best practices are ignored, creating an echo chamber
Sub-Methods 3
The smallest set of terms needed to ask good questions in a domain. For any new domain, there is a vocabulary threshold below which collaboration is inefficient and above which returns diminish. Identify this threshold: typically 20-30 terms that cover the core concepts. Invest in reaching it before starting production work.
Build something, hit a wall, name the wall, learn the vocabulary for it, continue. This is the natural learning loop in AI collaboration: the wall is the teacher. Each failure point reveals a concept you need to name. Document the vocabulary you acquire through doing so it persists across sessions and team members.
When you are iterating without progress, cycling through attempts that do not converge, diagnose whether you are missing vocabulary rather than missing skill. The signal is unmistakable: you revert, retry with different words, and get the same wrong result. The fix is to identify and name the concept you cannot currently articulate.