feat: add a "Verify before merging" checklist to the integration report

[ethangui]

Problem

Part of PostHog/wizard#447. After the wizard integrates PostHog there's a gap between "done" and "merged" — verification + glue steps the developer still needs. This adds them as a checklist the agent writes into posthog-setup-report.md.

Companion to the wizard PR PostHog/wizard#595, which surfaces a copy-paste handoff prompt pointing at this checklist.

Changes

llm-prompts/basic-integration/1.3-conclude.md now instructs the agent to write a ## Verify before merging checklist into the report, choosing only items relevant to the run from a curated candidate list.

Why this shape:

• The model picks the relevant items; the wizard doesn't compute them. Rather than a deterministic, wizard-side per-integration map, the agent judges each candidate against the code it actually changed this run.
• Curated candidate list + a "Do not invent items" guardrail — judgment within bounds, so it doesn't hallucinate steps.
• Strategy-agnostic wording — e.g. the LLM item is "trigger the LLM call path(s) you instrumented and confirm $ai_generation…", not assuming OpenTelemetry vs the @posthog/ai wrapper.
• It lives in the report, not the wizard — this is integration knowledge, so it belongs in context-mill, decoupled from the wizard release cycle.

Test plan

Validated via real wizard runs (real OAuth + agent + PostHog project) against a Next.js 15 SaaS app:

• the checklist rendered with the run-relevant items and omitted the rest;
• across runs, $ai_generation appeared only when LLM analytics was set up in that run and was correctly absent otherwise;
• confirmed the edited skill flows through the context-mill build into the served bundle.

Opening per the repo's contribution invitation in the README. Co-authored with Claude Code (Opus 4.8); follows the PostHog AI policy.

cc @PostHog/team-docs-wizard

🤖 Generated with Claude Code

[gewenyu99]

Amazing