mediar-ai Project Development Methodology
Identify tasks suited to LLM processing, design pipelines, and iterate with agent-assisted development.
Prerequisites
- Understanding of LLM capabilities
- Familiarity with batch processing
Instructions
Task-Model Fit
LLM-suited tasks:
- Synthesis across sources
- Subjective judgment with rubrics
- Natural language output
- Error tolerance acceptable
- Batch processing (no conversational state)
LLM-unsuited tasks:
- Precise computation
- Real-time requirements
- Perfect accuracy required
- Sequential dependencies
- Deterministic output needed
Manual Prototype First
Before building automation:
- Copy one representative input into model
- Evaluate output quality
- Identify failure modes
- Estimate tokens per item
If manual prototype fails, automated system will fail.
Pipeline Architecture
acquire → prepare → process → parse → render
- acquire: Fetch raw data
- prepare: Transform to prompt format
- process: LLM calls (expensive, non-deterministic)
- parse: Extract structured data
- render: Generate final outputs
File System as State Machine
data/{id}/
├── raw.json # acquire complete
├── prompt.md # prepare complete
├── response.md # process complete
├── parsed.json # parse complete
Check file existence to determine processing state.
Architectural Reduction
Start minimal. Vercel's d0 agent: 17 tools → 2 tools (bash + SQL), 80% → 100% success rate.
When reduction works:
- Data layer is well-documented
- Model has sufficient reasoning
- Specialized tools were constraining
Guidelines
- Validate task-model fit with manual prototype
- Structure pipelines as discrete, idempotent stages
- Use file system for state management
- Design prompts for parseable outputs
- Start minimal; add complexity only when proven necessary
- Estimate costs early and track throughout
Notes
- Karpathy's HN Capsule: 930 items, $58 cost, 1 hour
- Expect multiple architectural iterations
- Test whether scaffolding helps or constrains model
Source: muratcankoylan/Agent-Skills-for-Context-Engineering