| name | fresh-guides-verify |
| description | Invoked automatically when discussing a technology on the fresh-guides watchlist. Provides verification procedure: detect version, fetch official docs for that version, compare with training data, cite sources. Do NOT give version-specific advice for watched technologies without consulting this skill first. Keywords: fresh-guides verify, check docs, fast-changing technology, outdated knowledge, version check.
|
Fresh Guides — Verification Procedure
A technology on the user's fresh-guides watchlist was detected in the conversation. Follow this procedure BEFORE giving version-specific advice.
Steps
1. Identify the matched technology
Find the matching entry in the watchlist (injected by SessionStart). Note:
- The technology
name from the watchlist
- The configured
docs URLs
2. Determine the current version
Before giving any advice, establish which version of the technology is in use:
- Check the project — look for version pins in lockfiles, config files, or dependency manifests (e.g.,
terraform { required_version }, package.json, go.mod, .tool-versions, Dockerfile FROM tags, provider version constraints).
- Check runtime — if the tool is available locally, run its version command (e.g.,
terraform version, kubectl version --client, aws --version).
- If neither works — ask the user: "Which version of are you using?" Do NOT assume "latest" — the user may be on an older version with different behavior.
All subsequent verification MUST target the specific version identified in this step.
3. Determine what needs verification
Identify the specific claim, API, behavior, config syntax, or default that could have changed since training cutoff. Not everything needs verification:
Verify: API signatures, default values, config syntax, deprecations, new features, pricing, limits, CLI flags, provider arguments, service quotas.
Skip verification: General concepts, design patterns, architectural principles, well-established fundamentals that don't change across versions.
4. Fetch official documentation
Use WebFetch on the configured doc URLs to find information about the specific topic:
- Try the most specific doc URL first (e.g., API reference > general docs)
- When the official docs have versioned URLs (e.g.,
/v1.6.x/, /latest/), target the URL for the user's specific version — not "latest"
- If the page is too large or doesn't contain the answer, use
WebSearch with a targeted query:
"<technology> <version> <specific feature/API> official documentation"
- For release notes / changelogs, search for the specific version or recent changes
Fallback chain: configured docs URL → WebSearch official docs → WebSearch general → state "could not verify"
5. Compare with user's version and respond
After verification, explicitly compare the feature/behavior with the user's version:
- Feature available in user's version: State it clearly: "Supported in your version (vX.Y.Z, added in vA.B)."
- Feature NOT available in user's version: Warn explicitly: "Requires vA.B+. Your vX.Y.Z does not support it." Suggest a workaround if one exists for their version.
- Behavior changed between versions: Note what differs: "In your vX.Y.Z, the default is A. In vA.B+ it changed to B."
Based on what you found from docs:
Docs confirm your knowledge:
Include the answer with inline citation.
Docs contradict your knowledge:
Note: My training data may be outdated here. According to the official docs: [corrected information]. [Source: url]
Could not verify:
Note: I could not verify this against official docs. My answer is based on training data. Please double-check at: [configured doc URLs]
6. Citation format
Always include inline citations when providing verified information:
[Source: domain.com/path/to/versioned/page]
Place the citation after the specific claim it supports, not at the end of the entire response. Use version-specific URLs when the documentation site supports them.
Important
- NEVER silently use training data for version-specific advice on watched technologies
- NEVER skip verification because "I'm fairly confident" — the whole point is that confidence can be misplaced for fast-changing tech
- If WebFetch/WebSearch fails, say so explicitly — do not fall back to training data silently
- Keep verification efficient — fetch only what's needed, don't read entire documentation sites