session-guidance_skill

Overview

This skill configures Ralph loop session guidance using an interactive AskUserQuestion flow. It loads constraint scan results, presents dynamic forbidden/encouraged options based on severity, and writes validated guidance back to the project config for future sessions. It also records acknowledged constraints so they are filtered from future displays.

How this skill works

On invocation the skill checks for existing guidance and offers to keep or reconfigure it. It parses an NDJSON constraint scan, builds up to four dynamic forbidden options (critical → high → static fallbacks), prompts for custom forbidden and encouraged items, then validates and writes the combined guidance into the project config. Selected constraint IDs are appended to an acknowledgment store so they are excluded from future prompts.

When to use it

• When starting or restarting a Ralph session and you want to set or update forbidden/encouraged guidance
• After running a constraint scan and you need to convert findings into runtime guidance
• When the project mentions "forbidden items" or you need to prevent the agent from specific actions
• To persist learned exclusions so repeated findings are filtered from subsequent guidance prompts
• When you want a quick interactive reconfiguration without manually editing config JSON

Best practices

• Run the constraint scan step before invoking this skill so dynamic options reflect current issues
• Accept the recommended keep option when existing guidance already matches your policy to avoid duplication
• Prefer constraint-derived forbidden items for high-severity findings and use static categories to fill gaps
• Provide comma-separated custom items only when necessary; keep phrases concise for reliable matching
• Validate the written config after the skill completes and keep backups of prior config files

Example use cases

• A developer sees critical hardcoded secrets in a scan and selects them as forbidden so Ralph avoids touching those files
• A repo maintainer fills remaining slots with static categories like "Dependency upgrades" to guide automation
• A team reconfigures guidance before a release to prioritize "ROADMAP P0 items" and forbid refactors
• Automatically appending acknowledged constraint IDs so recurring low-risk warnings no longer clutter future prompts