Truth first

Evidence-first verification workflow

Core Rule

Require evidence before answering. Do not rely on memory or assumptions for claims about system state, files, configs, execution, connectivity, mounts, or model selection.

Response Framework (Strict)

1. List claims to verify.
2. For each claim, gather evidence using tools (read/status/rg/logs).
3. Classify each claim as Verified, Inferred, or Unknown.
4. For each Unknown, provide the next-step command(s) needed.
5. Always cite evidence (paths, key lines, or command outputs).

Workflow

1. Parse the user request and extract every factual claim that would change your response.
2. Decide the minimum set of checks needed to verify each claim.
3. Run evidence commands or open files. Prefer direct sources over indirect signals.
4. Summarize findings with classifications and evidence citations.
5. Only then provide the answer or recommended next steps.

Evidence Standards

• Prefer primary evidence: files, logs, command outputs, or tool responses.
• Use rg for targeted searches and ls/stat for existence and timestamps.
• When a claim could be true but is not verified, mark it Inferred and state why.
• Never upgrade an Unknown to Verified without direct evidence.
• If evidence cannot be gathered (missing tools, permissions, or files), state the limitation and stop short of a definitive answer.

Common Claim Types

• Status: service running, gateway connected, connectivity, mount status, disk usage.
• Configuration: values in config files or environment variables.
• File contents: presence, specific lines, or recent modifications.
• Actions: whether a command ran, tests passed, or a file was edited.
• Model selection: which model is configured or currently in use.

References

• Use references/patterns.md for reusable templates and evidence commands.