--- name: pd-intel-deep-dive description: Use when asked what people are saying about a post, topic, or person in PD Intelligence data — reaction analysis, sentiment, narratives, controversy, "how did X land", comment analysis, or finding who is driving a conversation. Chains post lookup in SQL, comment summaries, threads, and commenter profiles. --- # PD Intelligence — Post & Comment Deep-Dive This answers in chat. When the user wants an artifact someone else reads — a document, a deck, an interactive story — gather the evidence here, then render it with `pd-intel-report`. Answer "what are people saying and who is saying it" with evidence: actual quotes, counts, and ids — not vibes. ## Prerequisites - Call `list_datasets` first; pass the dataset's integer `id` everywhere. - **Comment text is not collected for every platform** (historically not X/Twitter, even though posts there show a `comments` count). Check `get_comment_summary`'s `platform_stats` for actual coverage, and say "comments aren't collected for this platform" instead of "no discussion". - Posts have two ids: internal `id` (use this; it is the `post_id` the comment tools take and return) and the platform's `platform_post_id` (the id in the post URL, in the SQL `posts` table). ## Workflow 1. **Locate the subject.** `run_analytics_sql` over `posts` JOIN `post_metrics_latest` (`ON post_metrics_latest.post_id = posts.id`), with `post_text ILIKE '%keyword%'` (or `title`), `author_username`, or `platform_post_id` (the id from a post URL). `author_username` is the post's byline: on a collab it can be the partner, not the tracked account. For one tracked account's posts, pass its id as `account_ids` (from `search_accounts` or the `accounts` table) instead. Pass the window and platforms as the `date_from`/`date_to`/`platform` **arguments**, not a `WHERE`. For a topic, run a `count(*)` first — that's the conversation's size — then fetch the top rows by views with `post_text` named explicitly. Patterns: the navigation skill's `references/analytics-sql.md`. 2. **Post facts**: the same join with `WHERE p.id = …` for the focal post(s) — metrics, `post_url`, `content_type` — plus their tags from `entity_tags` (`entity_type = 'post'`; `is_direct` separates direct tags from ones inherited from the account or creator). 3. **Reaction shape**: `get_comment_summary` with `post_id` for one post, or date-filtered for a topic/period. Gives volume, unique commenters, author-reply rate, timeline, and the most-discussed posts. These count the comments PD holds, which can be far fewer than the post's own `comments` metric (collected once, about three days after the post). Its `capture` block gives the share held per platform; state both numbers and that share, and say how many held comments you read, before generalizing. 4. **Read actual comments**: `search_comments` with `post_id` (or `search` keywords), two passes: - `sort_by: "likes"` — what the audience endorses (top comments). - `sort_by: "comment_timestamp"` — how the reaction is evolving. Sample 25-50 before characterizing sentiment; quote 3-5 representative comments verbatim with their like counts. 5. **Threads, carefully**: `search_comments` with `root_comment_id` returns the root plus every reply, oldest first — but a popular comment can have hundreds of replies. Check `replies_count` first (`sort_by: "replies_count"` lists the biggest threads); only page through threads that are small or genuinely pivotal. `is_author_reply: true` in `search_comments` finds where the creator engaged. 6. **Who's driving it**: for recurring usernames, `list_commenters` with `commenter_id` (from the comment) gives their profile row and `total_comments`, and `search_comments` with the same `commenter_id` shows their history; `list_commenters` sorted by `total_comments` surfaces the dataset's loudest voices. Note `platforms_active` — cross-platform commenters are notable amplifiers. For one creator's or account's repeat commenters, `get_audience` (`creator_id` or `account_id`, dates) returns the superfans, a bot-like share and a `core_verdict` (`organic` / `orchestrated` / `too_few`) in one call — use it before calling a reaction coordinated. For voices that span several of the dataset's accounts, `get_cross_account_commenters` (tag-scoped, `min_accounts`, dates) finds them in one call instead of many `search_comments` passes. 7. **What changed in the conversation**: `get_rising_terms` with the period as `date_from`/`date_to` (tag and `platform` filters optional) compares word and hashtag use against the previous period of the same length — `rising` and `falling` terms with `comments_now`, `comments_before`, `ratio` and two example comments each. It is computed in the warehouse, so it beats keyword counts in SQL on a busy dataset. Read `note`: a range ending in the last three days is still filling in. ## Analysis standards - **Separate volume from valence**: report how many commented, then what they said. High likes on critical comments ≠ overall negativity — weigh like counts, not just comment counts. - **Name narratives, then prove them**: each claimed theme needs at least two quoted examples with comment ids or usernames. - **Watch for coordinated/repeat voices**: same commenter across many posts, near-identical text, or single-issue accounts — check with `list_commenters(commenter_id=…)` and `search_comments(commenter_id=…)` before treating them as organic sentiment. - **State sample size**: "of the top 50 comments by likes…" — never imply you read a million comments. - **State the denominator** before calling a theme common: how many comments you read, and how many carried it. - **A missing metric is not a finding.** Bluesky publishes no view count and Instagram publishes none on photos or carousels, so a post's `0` views says nothing about how the post landed. Judge reach only on platforms that publish it, and never write a caveat about "missing data" for a metric a platform simply never publishes — that is method, not a limitation.