alighasami
3d5eaf9445
Security: Sync from Public / sync-from-public (push) Has been cancelled
Test: Benchmark Nightly / build (push) Has been cancelled
Test: Benchmark Nightly / Notify Cats on failure (push) Has been cancelled
CI: Python / Checks (push) Has been cancelled
Test: Evals Python / Workflow Comparison Python (push) Has been cancelled
Util: Check Docs URLs / check-docs-urls (push) Has been cancelled
Test: Visual Storybook / Cloudflare Pages (push) Has been cancelled
Test: E2E Performance / build-and-test-performance (push) Has been cancelled
Test: Workflows Nightly / Run Workflow Tests (push) Has been cancelled
Util: Cleanup CI Docker Images / Delete stale CI images (push) Has been cancelled
Test: Benchmark Destroy Env / build (push) Has been cancelled
Util: Update Node Popularity / update-popularity (push) Has been cancelled
Test: E2E Coverage Weekly / Coverage Tests (push) Has been cancelled
1.3 KiB
1.3 KiB
AGENTS.md
Guidance for working with the AI Workflow Builder package.
Prompt Development
Use the PromptBuilder utility for all LLM prompts. See src/prompts/README.md
for detailed documentation.
import { prompt } from '@/prompts/builder';
const systemPrompt = prompt()
.section('role', 'You are an assistant')
.sectionIf(hasContext, 'context', () => buildContext())
.build();
Key methods:
section(name, content)- Always includedsectionIf(condition, name, content)- Conditionally includedexamples(name, items, formatter)- Formatted example listbuild()- Returns the final prompt string
Workflow Examples in Prompts
When including workflow examples in prompts, use Mermaid flowcharts instead of raw JSON. Generate Mermaid from workflow JSON using:
pnpm workflow:mermaid path/to/workflow.json
This outputs a markdown file with a Mermaid diagram. The diagram is more readable for the LLM and uses fewer tokens than JSON.
Options:
--no-node-name- Exclude node names--no-node-type- Exclude node type comments--node-params- Include node parameters
Evaluations
Before modifying evaluations, judgement criteria, or scoring logic, read
evaluations/README.md for the evaluation framework architecture and guidelines.