Training memory drifts. Wrong assumptions compound across turns. This is a quick source check, not full deep research.
When this applies
External systems: CLIs, install flows, API behavior, registry or publish claims, version-specific flags, third-party tool behavior.
Skip for: code already in the workspace, pure opinion, or conversation recap macros.
Process
- Name the claim — one sentence you were about to assert.
- Verify —
Readin-repo docs orWebFetch/ search official source once. Do not chain guesses. - Answer with a label:
- verified — cite URL or file path (one line)
- unverified — say what you tried; do not state as fact
- memory — not checked — only if the user explicitly asked to skip lookup
Output shape
Reply with only:
claim
source check
<what was read or fetched, or "none — unverified">
answer
<corrected statement, or honest "don't know yet">
label
verified | unverified | memory — not checked
Hard rules
- Do not invent CLI flags, install steps, or directory layouts from memory.
- If docs conflict with memory, docs win.
- For ledger-grade research with confidence scoring, use the
verasic-deep-researchskill instead.