align

Adjust system prompts and gate logic to align website-agent builds with Braintrust traces.

Updated Apr 17, 2026

Search Tags:#e2e #traces #braintrust #system-prompt #derailment #align

Ready to run

npx skills add https://github.com/Aradotso/ara.engineer --skill align-aradotso

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: align
Source: https://github.com/Aradotso/ara.engineer/tree/main/skills/align
Command: npx skills add https://github.com/Aradotso/ara.engineer --skill align-aradotso

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Align the website-agent with Braintrust traces to reduce derailments and ensure builds progress smoothly by adjusting system prompts and gating rules. The skill provides a structured approach to diagnose derailments, patch prompts, and verify improvements through a tight loop of traces and e2e validations.

Core Features & Use Cases

Trace-driven alignment: fetch traces, identify derailment patterns, and map them to prompt and gate adjustments.
Iterative patching: modify system-prompt.ts and related server logic to fix wrong phase or tutoring drift.
Verification loop: run e2e tests, replay traces, and report before/after fit with scores from Braintrust evaluators.

Quick Start

Patch the system-prompt and related server logic to fix derailments, then run the local e2e replay to verify improved fit.

Frequently Asked Questions about align

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I fix system prompt derailments using Braintrust traces?▼

System prompt derailments can be fixed by fetching Braintrust traces, identifying drift patterns, and patching system-prompt.ts and server logic. This trace-driven alignment reduces unexpected agent behavior during site builds.

What is the best way to align an AI agent with end-to-end test evaluations?▼

Aligning an AI agent with e2e evaluations involves modifying backend system prompts and gate logic, then running e2e replay tests to verify improvements. Braintrust evaluators provide before and after fit scores for validation.

How does trace-driven prompt patching work for website builds?▼

Trace-driven prompt patching works by analyzing Braintrust traces to detect phase errors or tutoring drift, modifying the system prompt to correct the logic, and validating the changes through a fast e2e feedback loop.

Can I use e2e tests to validate system prompt changes in a local environment?▼

Yes, you can validate system prompt changes locally by running e2e replay tests. This verification loop replays traces to report before and after fit, ensuring prompt patches reduce derailments effectively.

Why does my website agent enter the wrong phase during automated builds?▼

Website agents enter wrong phases during builds due to system prompt drift or inadequate gating rules. Adjusting the system-prompt.ts file and related server logic based on trace analysis corrects this behavior.

scooter-lacroix

braintrust-tracing

Capture Braintrust traces for Maestro AI sessions and correlate sub-agent executions.

Community

Advanced

Aradotso

replay-railway-preview

Deploys nightly build to preview and executes timed test suite with health checks and automated cleanup, generating a final report and optionally triggering CI/CD pipeline.

Official

Advanced