An open-source repository called Salesforce Ada Agent Skills introduces a structured '/decide' pipeline that lets AI coding agents like Claude Code, Gemini, and Codex CLI help make and document technical decisions. The skill walks through requirement gathering, option research against official documentation, criteria definition, risk scoring (low/medium/high mapped to fixed numeric values), and a weighted decision score based on Salesforce's Well-Architected framework dimensions (reliability, resilience, operational excellence, resource optimization, cost optimization, and equality). Human review checkpoints occur after each step, and the final recommendation is written to disk as a Markdown Architectural Decision Record for future reference. The design separates pipeline logic (SKILL.md), rules and weights (CLAUDE.md), and output templates into different files so the agent's context stays focused and formats can be iterated independently. It works with or without a documentation-search MCP server, though a custom MCP server indexing official Salesforce docs is recommended for higher-quality sourcing.
Questions this post answers
How does the /decide skill calculate a weighted decision score for technical options in AI coding agents?
Risk levels are mapped to fixed confidence numbers (low risk = 85, medium = 55, high = 25), then multiplied by weights for six evaluation axes based on the Well-Architected framework: reliability 20%, resilience 20%, operational excellence 20%, resource optimization 15%, cost optimization 15%, and equality 10%. The option with the highest weighted total becomes the recommendation, with no room for the agent to override the score through judgment. Teams designing agent-driven decision workflows can track patterns like this through daily.dev.
Why does the decision-scoring skill keep the scoring rubric in a separate file loaded only at the final step?
The scoring methodology file contains roughly 350 lines of detailed evaluation criteria, and loading it early would distract the agent's attention during the information-gathering phase. Splitting files this way lets teams tweak formatting or scoring weights without touching the pipeline logic itself, keeping each step's context minimal and focused. Developers structuring multi-step agent skills can compare approaches like this on daily.dev.
Which AI coding agents can run the Salesforce Ada Agent Skills /decide pipeline?
The skill is demonstrated using Claude Code but is designed to work with any AI agent, with templates also provided in the repository for Gemini and Codex CLI. It can function without a documentation-search MCP server by falling back to web search or the agent's trained knowledge, though a custom MCP server indexing official Salesforce docs is recommended for higher-quality results. Anyone evaluating agent-agnostic skill designs can follow ecosystem updates like this via daily.dev.
Share this post