truth-frist_skill

Overview

This skill enforces an evidence-first approach before answering claims about system state, files, configs, actions, connectivity, mounts, or model selection. It prevents assumptions by requiring concrete verification steps and documented evidence. Use it as a gating layer before any diagnostic, remediation, or configuration response.

How this skill works

The skill parses a request and extracts every factual claim that could affect the answer. For each claim it runs the minimal set of evidence-gathering commands (file reads, status checks, logs, searches) and classifies results as Verified, Inferred, or Unknown, citing paths, output lines, or timestamps. For Unknown items it returns exact next-step commands and stops short of definitive conclusions until evidence is available.

When to use it

• Before answering questions about running services, mounts, network connectivity, or disk usage.
• When asked to confirm configuration values or environment variables that affect behavior.
• Before asserting that a file contains specific content or was recently modified.
• Prior to claiming that an action was executed, tests passed, or a deployment completed.
• When model selection or runtime model identity must be known before responding.

Best practices

• Always enumerate explicit claims the user expects verified before running checks.
• Prefer primary evidence: read files, run status commands, and cite exact lines or timestamps.
• Use targeted searches (rg) and ls/stat for existence and modification time checks.
• Classify results strictly: Verified only with direct evidence, Inferred with reasoning, Unknown otherwise.
• For Unknown items, provide precise next-step commands and do not guess.

Example use cases

• Verify whether a service is active and its last log lines before suggesting a restart command.
• Confirm a configuration setting in a specific config file and show the matching line before recommending changes.
• Check that a mount is present and list the mountpoint's contents before advising on disk operations.
• Inspect which model binary or endpoint is configured and provide the command outputs that prove it.
• Search repo files for a referenced commit or change and cite the exact file path and line.