# Validation and practical limits

Checked 4 November 2026.

## Package

- Skill frontmatter/name validated with the skill-creator validator. Description is 180 characters; package has only a portable `SKILL.md` with name/description metadata.
- The pasteable prompt contains the identical workflow body, with an input brief added.
- `customer-question-research.zip` contains exactly `customer-question-research/SKILL.md`, with matching file bytes. No scripts, dependencies, private notes or external connections.
- `attendee-kit.zip` contains the guide, pasteable prompt, examples, this validation note and the skill ZIP. **Upload the inner `customer-question-research.zip` to Claude, not the entire attendee kit.**

## Fresh Claude Code check

A temporary project held a copy at `.claude/skills/customer-question-research/SKILL.md`. A fresh Claude Code **2.1.218** invocation listed `customer-question-research` in its initialization event, confirming discovery of this folder/name. The request attempted slash invocation with read-only research tools, no MCP servers and hooks disabled.

The session then returned **“Failed to authenticate: OAuth session expired and could not be refreshed.”** No model response or research tool call followed. Discovery was observed; Claude execution quality, slash expansion and browsing behavior were **not** verified. No account changes, global installation or publishing were performed.

## Fresh prompt-workflow check

A separate fresh **Codex** session received the exact skill workflow and a bounded public-source pack. It used supplied evidence only, with no browsing, tools or file changes. This tests instruction behavior in another model; it is not a substitute for a Claude model test.

Isolation was audited before public distribution: a new temporary working directory, new ephemeral CLI process, no `resume`/`fork`, no prior conversation, expected answer or previous model output in its input. User configuration/rules were not loaded. The exact workflow, sample request and public pack were the only supplied task context. No rerun was needed. The local audit is retained in `validation/isolation-audit.md`.

Observed result:

- Returned three ranked topics and two draft hooks per topic, with evidence IDs and source links.
- Kept customer statements, interpretation and competitor messaging separate.
- Excluded the deliberately search-only C3 lead from substantive support and recurrence counts; did not imply it had independently browsed.
- Left the unknown business's offer facts unverified, including supplies, attendance rules and continuity.
- Recognized that independent evidence is needed for trust and stopped before scripting/production.

The original test pack and model output are retained in the local `validation/` folder for review; they are excluded from the attendee kit. The attendee worked example is an edited illustration separately checked against opened public pages, including C3 after the test pack was compiled. It therefore labels C3 as verified, while the unchanged test output appropriately excludes it.

## Still untested

Claude web ZIP upload, enable/automatic activation, Claude model output, Claude web browsing and cross-account sync. The installation instructions follow current official documentation, linked in the guide. This machine's older Claude Code version does not support the documented newer account-sync route. No behavior is claimed merely from packaging or documentation.