session-intent

The idea: you own the page's source, so you know what each visitor was supposed to do. Write that down as an ordered intent spec, then score every PostHog session against it. The output is one finding: the step people fail at, the counts, two replay links, and the fix. Under 200 words.

The pipeline (build it as a small script)

1. List hosts/pages with sessions (HogQL over events grouped by $host, $pathname).
2. Pull sessions for one page over N days, with their events and, where they exist, the recording snapshots.
3. Score each session step by step against the intent spec (below): reached, missed, or unknown.
4. Summarise with counts computed in code, then let a model write the finding from the counts and a few exemplar sessions. The model never writes a number: counts come from the scorer and are printed by code.

Writing an intent spec (sites/<slug>/intent.json)

Derive it from the SOURCE; don't ask the site owner.

1. Find the template or generator, not the built HTML. Pin source_commit with git rev-parse --short HEAD.
2. job: one sentence, in the visitor's words, naming who it was made for.
3. Steps in order: land -> see the thing -> (optional narrow) -> take it / act on it. success is the step that means the job got done. Each step's source names file:line, and includes the layout CSS lines too (the grid and its media query). Without them a model will propose a DOM move that breaks desktop.
4. match alternatives per step. Find which custom events the page sends: grep send(' / posthog.capture and any shared tracker. Kinds: pageview, event (+prop, min / equals / contains), scroll_px (rrweb scrollY), click (substring of tag#id.class "text"), input.
5. Measure positions at the visitor's real viewport with headless WebKit (e.g. 393x659): first row top, CTA top, document height. Put them in measured_at. A scroll step's scroll_px threshold = element top minus one viewport, plus a little.

{"slug": "pricing-page", "url": "https://example.com/pricing/", "source_commit": "abc1234",
 "job": "A shop owner checks what the plan costs and starts a trial.",
 "measured_at": {"viewport": "393x659", "cta_top": 1420, "doc_height": 3900},
 "steps": [
  {"id": "land", "match": [{"kind": "pageview"}], "source": "templates/pricing.html:1"},
  {"id": "see_price", "match": [{"kind": "scroll_px", "min": 800}], "source": "templates/pricing.html:42, static/site.css:118-131"},
  {"id": "start_trial", "success": true, "match": [{"kind": "event", "name": "trial_click"}, {"kind": "click", "contains": "Start trial"}], "source": "templates/pricing.html:77"}
 ]}

Rules (each one cost a cycle)

• Scroll % is meaningless on long pages. On a 100k-px list, 3,785px scrolled reads as max_depth 0. Use scroll_px or an in-list interaction with "decides_miss": false. Never use a % event alone.
• filterTestAccounts does not catch QA bots. Drop bursts (3+ distinct ids, one device profile, within 120s) and print a left out: line before quoting a count.
• Check the device count before calling it a crowd. 18 sessions can be one iPhone reopening a restored Safari tab (every replay opens at the same scrollY and hides in 0.2s). Never guess who a visitor is; describe the device and the pattern.
• Verify the model's fix before reporting it: inject the CSS or JS into the live page with Playwright add_style_tag at 393 AND 1440 width, and measure. Verify its tap claims against the cached session JSON. Models tend to understate counts.
• Most sessions may have no recording (lazy-loaded recording). Events-only steps are scored unknown unless an instrumented event can decide them.
• Never edit the site from this skill without the owner's word. Report the fix.

PostHog API traps

• Snapshots: GET /api/projects/<id>/session_recordings/<sid>/snapshots?blob_v2=true lists sources, then fetch each with source=blob_v2&start_blob_key=k&end_blob_key=k (both are required, otherwise 400). JSONL of [window_id, event]. Payloads with cv: "2024-10" are gzip carried as latin-1 strings. Split on \n only: splitlines() breaks inside them.
• HogQL filterTestAccounts: pass "filters": {"filterTestAccounts": true} AND put {filters} in the WHERE clause, or it does nothing. A shell eats $session_id, so query from Python, not curl.
• /session_recordings/<sid> 404 means no recording, even when the session has events.
• Use a personal API key with read scopes only.