fix(seo): og:image jpg not webp so LinkedIn renders social previews

docs/projects/2605-tech-for-non-technical-founders/TASK-TRACKER.md

@@ -10,7 +10,7 @@
 
 ## Open queue (in execution order)
 
-0. **Sprint "Prove ICP Arrival" (validation) — ACTIVE, set 2026-08-13.** Manager panel (growth + learning + lean lenses, `/brainstorm`+vote) UNANIMOUS: with no traffic lift (84→75 sessions/14d, ~93% Direct, course absent from GSC) AND a dead funnel (start_course=1, mechanics=0), the constraint is **arrival/reach, not polish**. **Audit correction (premise was stale):** the campaign for this is ALREADY BUILT and READY — `linkedin-posts/icp-validation/POSTING-PACKET.md`, 10 posts drafted, zero posted, UTM scheme defined (`utm_campaign=icp_validation_<pillar>`), **▶ POST #1 SCHEDULED 2026-08-13** on Paul's LinkedIn for Thu Aug 13 3:30 PM CEST (9:30am ET peak) via claude-in-chrome, verified in Scheduled posts — 2-week kill-clock starts on publish today. Post #1 is a **reply-CTA** post (validates ICP *presence*/language on LinkedIn, H1-H5), which is the correct **cheapest FIRST test** — upstream of course *arrival*. **Step 2 (backlog, manager-prepped): UTM'd course-link posts** to test click-through/arrival, drafted only if step-1 engagement appears (contradicts H5's no-link-CTA design, so gated on step-1 signal). **Manager autonomous work while Paul-gated:** verify packet still valid post-Aug-14 (✅ story post evergreen, passed 5 critic passes), prep step-2 arrival posts + the 5-day measurement/kill-criteria read. **Panel parked, data-justified:** rest of W4 (item 13) dead-mechanic removal → opportunistic; 36-image SVG migration → opportunistic; item 7 → BLOCKED by gate #1.
+0. **Sprint "Prove ICP Arrival" (validation) — ACTIVE, set 2026-08-13.** Manager panel (growth + learning + lean lenses, `/brainstorm`+vote) UNANIMOUS: with no traffic lift (84→75 sessions/14d, ~93% Direct, course absent from GSC) AND a dead funnel (start_course=1, mechanics=0), the constraint is **arrival/reach, not polish**. **Audit correction (premise was stale):** the campaign for this is ALREADY BUILT and READY — `linkedin-posts/icp-validation/POSTING-PACKET.md`, 10 posts drafted, zero posted, UTM scheme defined (`utm_campaign=icp_validation_<pillar>`), **▶ 2026-08-13 UPDATE:** the rescue-lane jira post was scheduled then PULLED (Paul: AI slop). It exposed the pipeline's broken "story, not advice" doctrine → corrected to **idea-first, deliver the point** (`linkedin-post-pipeline.md` + new `linkedin-posts/README.md` LI operating system: 2 lanes course/rescue, 2-3 posts/wk, frontmatter v2, `metrics-ledger.md`, `reference-examples.md` from Cutler/Rossi/Holub/Paolino). First idea-first post SHIPPED: course-lane `course-promo/week1-thu-validate-before-build.md`, scheduled 3:30pm CEST (text-only; image = manual gap). **Next wave: revise the ~10 existing lane drafts to the new doctrine, 2-3/wk, Paul-gated.** Post #1 is a **reply-CTA** post (validates ICP *presence*/language on LinkedIn, H1-H5), which is the correct **cheapest FIRST test** — upstream of course *arrival*. **Step 2 (backlog, manager-prepped): UTM'd course-link posts** to test click-through/arrival, drafted only if step-1 engagement appears (contradicts H5's no-link-CTA design, so gated on step-1 signal). **Manager autonomous work while Paul-gated:** verify packet still valid post-Aug-14 (✅ story post evergreen, passed 5 critic passes), prep step-2 arrival posts + the 5-day measurement/kill-criteria read. **Panel parked, data-justified:** rest of W4 (item 13) dead-mechanic removal → opportunistic; 36-image SVG migration → opportunistic; item 7 → BLOCKED by gate #1.
 
 1. **Item 16 — Landing full-migration to the shuffle2 reference. ✅ DONE, MERGED (PR #444, squash `902686f3`, 2026-08-12).** Paul approved the render + flagged `bin/check-landing-parity` as redundant → removed (YAGNI). Design panel (conversion/taste/UX) decided Fork A = Option 2 + embedded CTA (dark NOT-cover, endcap merged into the dark base, no light sliver), Fork B = relocate all 4 off-reference sections, section reorder to reference order. Shipped T1-T6 (relocate 4 sections + 5 anchor repoints, reorder, gradient H1, 3-col mistakes, hero stat cells, dark band + merged endcap + 4 dead-selector cleanup, de-farmed chapter links). Linux baselines re-recorded via CI `update-baselines` (40-file env-drift re-sync incl. the intended `desktop/course/landing`) and rode the merge. Result render: `40-49-review/40.32-item16-result-full.jpeg`. Both reviewer agents failed on Fable quota → all blocking checks self-run with grep evidence. **Landing is live.**
 

docs/workflows/linkedin-post-pipeline.md

@@ -107,41 +107,52 @@ Every line of every LinkedIn post must pass:
 
 ---
 
-## Post shape: story, not advice (BLOCKING)
-
-The single biggest tell that a post is sales/marketing is its **shape**, not its words. Marketing posts follow a predictable arc: hook → anecdote → generalization → solution → objection-handling → empowerment → CTA. That shape itself reads as marketing even when every individual word is fine.
-
-Paul's posts must be **story-shaped**, not advice-shaped:
-
-| ❌ Advice shape (marketing) | ✅ Story shape (conversation) |
+## Post shape: idea-first, deliver the point (BLOCKING)
+
+> **Doctrine corrected 2026-08-13.** The prior rule ("story, not advice" - force
+> every post into a single-encounter parable) *generated* the AI slop Paul rejected:
+> staged "a founder pinged me last Tuesday" openers, the tactic buried as the
+> founder's future action (zero takeaway), and "So... So..." connectors. It is
+> replaced below. Reference writers Paul rates: **John Cutler** (`johnpcutler`) and
+> **Luca Rossi** (`lucaronin`, refactoring.fm) - study their recent activity.
+
+The marketing tell is the **arc**, not advice itself. Sales posts run hook → anecdote → generalization → solution → objection-handling → empowerment → CTA. Drop the arc - do NOT hide the point inside a parable.
+
+**Write idea-first, like Cutler and Rossi:**
+- **Open with the idea, flat.** A belief, observation, or claim - not a staged encounter. Rossi: *"One of the dominant narratives in software today is that AI is making junior engineers redundant."* Cutler: *"'We want visibility' is stated as a universal need. But..."*
+- **Argue it plainly, then take a position.** Rossi: *"It goes like this... I believe this is a mistake."*
+- **Deliver the tactic in full.** Advice IS the value - give the actual move away. (Opposite of the old "hide the tactic inside what the founder will do next" rule.)
+- **Plain, short sentences.** No stacked sub-clauses. Paul: "I write simpler."
+- **Paragraph breaks are fine.** Cutler and Rossi use clear one-idea paragraphs. The old "don't separate the beats, add So-connectors" rule is **REVOKED** - it manufactured the beat-marking Paul flags as slop.
+- **Close with a real peer question.**
+
+| ❌ Old parable shape (now banned) | ✅ Idea-first shape |
 |---|---|
-| Open with a tagline ("Jira is not progress.") | Open with a specific recent encounter ("Founder showed me her sprint board last week.") |
-| Generalize the pattern in the middle ("This shows up in almost every team I look at...") | Stay inside the encounter, let the reader generalize |
-| Lift the tactical advice into a how-to section ("What works is asking for a URL...") | Embed the tactical move inside what the founder is going to do next ("She's going to try something this Friday: ...") |
-| Add an objection-handling bullet list ("- We're refactoring → ...") | Skip the objection list. If the post must address objections, fold them into one sentence inside the story |
-| Close with empowerment + question ("Any non-technical founder can do this. What's your story?") | Close with a single peer question ("Anyone else been in this version of it?") |
+| Staged encounter opener ("A founder pinged me last Tuesday...") | The idea, flat ("A lot of non-technical founders read a full board as proof of progress.") |
+| Bloat the encounter into a 4-paragraph screenplay with "So... So..." connectors | Argue the pattern in plain sentences |
+| Bury the tactic as the founder's future action; deliver no advice | Deliver the tactic directly - reader can use it today |
+| Manufactured specificity (14 tickets, since January, this Friday, ten minutes) | One real detail if you have it, else none - don't fabricate a tidy composite |
 
-**The shape test:** Read the post in your head. If it could appear unchanged in a "5 ways to spot a stalled dev team" newsletter, the shape is wrong. Rewrite as a recounted encounter, not as advice with a story decorating it.
-
-### Story-shape skeleton
+### Idea-first skeleton
 
 ```
-Beat 1 (1-2 lines): Specific recent encounter - concrete detail (number, day, role)
-Beat 2 (1-2 lines): The reveal moment in the encounter
-Beat 3 (1-2 lines): One line of opinion as observation, not slogan
-Beat 4 (1-2 lines): What the person in the story is going to do next
-Close (1 line):     Peer question - "anyone else been in this version of it?"
+Beat 1 (1 line):    The idea / belief / observation, flat - no staged encounter, no credential
+Beat 2 (1-2 lines): Restate it plainly, or show why it looks reasonable
+Beat 3 (1-2 lines): Your position, flat
+Beat 4 (3-5 lines): The tactic, delivered in full - the actual move the reader can make
+Beat 5 (1 line):    Why it matters
+Close (1 line):     Real peer question
 ```
 
-Length target: **120-180 words**. Story shape needs less transition scaffolding than the marketing arc, so posts run shorter.
-
-### First-draft warning: don't separate the beats into paragraphs
-
-The skeleton names 5 beats. The first-draft trap is to render each beat as its own paragraph with no connector to the prior beat. That produces a list-of-points shape that reads as outline, not story — the post will fail the read-aloud fluency test even if every individual sentence is clean.
+Length target: **120-160 words**. Plain sentences, one idea per paragraph.
 
-**Fix at the writing stage:** drop in connectors at the seams as you draft. "So we hopped on a quick call." "And the thing is..." "Anyway, we agreed she'd try..." Real spoken stories don't pause for breath at every clean beat. Two paragraphs of flowing prose almost always beat 5 paragraphs of separated beats.
+### New slop tells (2026-08-13 - caught on the jira-not-progress + validate-before-build drafts)
 
-**The Tuesday post lesson:** an initial 6-paragraph draft (one paragraph per beat, no connectors) was rewritten to 3 paragraphs by adding `so we hopped on a quick call`, `So we agreed she'd try a different one`, and `Beats three more sprints of...` — connectors that carry the listener forward. This is a writing-stage rule, not just an editing-stage fix.
+- **Manufactured single-encounter parable** - a too-clean composite founder story built to illustrate a lesson.
+- **"So... So... So..." beat-marking** / cinematic connectors ("So we hopped on a call") - screenplay reconstruction. (The old doctrine *mandated* these.)
+- **Cost-stacked time-cut** - "six months and a big invoice later, you still..." - manufactured drama.
+- **Stacked-clause long sentences** - one 30-word sentence with 3 sub-clauses. Break it up.
+- **Engagement-bait close as the only payload** ("Anyone else been in this version of it?") when the body delivered no value. The question is fine *after* real value; not as a substitute for it.
 
 ### Hook archetype rotation (Beat 1)
 

layouts/partials/seo/enhanced-meta-tags.html

@@ -111,7 +111,8 @@
       {{- $resource := $page.Resources.Get . -}}
       {{- if $resource -}}
         {{- if site.Params.cdn.enabled -}}
-          {{- $image = partial "cdn/url" (dict "page" $page "resource" $resource "params" "w=1200&h=630&fit=cover&output=webp&q=85") -}}
+          {{/* og:image must be jpg/png - LinkedIn's crawler won't render WebP (Telegram/most do). Keep jpg for reliable social unfurl. */}}
+          {{- $image = partial "cdn/url" (dict "page" $page "resource" $resource "params" "w=1200&h=630&fit=cover&output=jpg&q=85") -}}
         {{- else -}}
           {{- $optimized := $resource.Resize "1200x630 jpg q90" -}}
           {{- $image = $optimized.Permalink -}}

linkedin-posts/README.md

@@ -0,0 +1,135 @@
+# LinkedIn Operating System — Paul Keen / JetThoughts
+
+How we run Paul's LinkedIn: two lanes, a slow cadence, one voice doctrine, and
+per-post analytics tracking. Nothing publishes without Paul.
+
+## Voice + shape (the engine)
+
+**Canonical:** `docs/workflows/linkedin-post-pipeline.md` → "Post shape: idea-first,
+deliver the point (BLOCKING)". Corrected 2026-08-13 away from the old "story, not
+advice" parable doctrine that produced AI slop. In one line: **open with the idea
+flat, argue it plainly, deliver the tactic in full, plain short sentences, close
+with a real peer question.** Study the writers Paul rates: John Cutler
+(`johnpcutler`), Luca Rossi (`lucaronin`/refactoring.fm). Good examples transcribed
+in [`reference-examples.md`](reference-examples.md).
+
+## The two lanes
+
+| Lane | Folder | Stage | ICP | Plan |
+|---|---|---|---|---|
+| **Course** (pre-validation) | `course-promo/` | idea-stage / about-to-build | validate demand *before* building | `docs/workflows/linkedin-course-promo-plan.md` |
+| **Rescue** (control-loss) | `icp-validation/` | already building, can't verify | founder stuck with a dev shop | `docs/workflows/linkedin-icp-validation-plan.md` |
+
+Both post from Paul's personal account. **Rotate between lanes** so the feed isn't
+all one stage. Assets (images) live in each lane's `assets/` folder.
+
+## Cadence (2026-08-13, Paul)
+
+**2-3 posts per week, max** — the best-for-ICP volume (non-technical founders don't
+want a daily firehose). Post at the ICP's peak read time: **Tue-Thu, US-morning
+(9-11am ET = ~15:00-17:00 CEST)**. Schedule via LinkedIn's native scheduler. One
+lane per post, alternating.
+
+## Every post carries a visual (Paul: "missed images")
+
+Each post gets one image in its lane's `assets/`. Reuse course exhibits (the
+refactoring.fm-style SVGs) exported to PNG via `rsvg-convert -w 1080 <svg> -o <png>`
+— they're on-brand and seed the course bridge without a link. Frontmatter records
+the path.
+
+> **Known tooling gap:** claude-in-chrome can schedule text but cannot drive
+> LinkedIn's native file-upload dialog. Until solved, the image is attached
+> manually by Paul (or dropped in a fresh composer). Track the fix in the backlog.
+
+## Frontmatter schema v2
+
+Lifecycle + pointers in frontmatter; **performance data lives in the ledger, not
+the frontmatter** (single source, no double-authoring — see below).
+
+```yaml
+---
+lane: course | rescue
+week: 1
+day: thursday
+author: paul-keen
+pillar: <e.g. demand-before-build | progress-visibility>
+hypothesis: <which validation hypothesis>
+opener_archetype: idea-led | observation-led | question-led | stat-led | conflict-led
+icp_test: <one line - what this post tests>
+visual: assets/<slug>.png        # required - the image that ships with the post
+status: draft | scheduled | posted
+scheduled_for: <ISO date, if scheduled>
+posted_url: <LinkedIn URL, once posted>
+first_comment: |            # posted as the FIRST COMMENT right after publish (not the body)
+  <course lane: the ready comment incl. the UTM'd course link. rescue lane: usually empty.>
+notes: | <voice trade-offs, revision history>
+---
+```
+
+## Link policy per lane (BLOCKING)
+
+- **Never a link in the post body** - LinkedIn throttles reach on external links, and JT voice bans in-body CTAs.
+- **Course lane** → one UTM'd course link in the **first comment** (`first_comment` in frontmatter). Link the specific lesson that delivers on the post's promise. This is the arrival signal the metrics-ledger tracks.
+- **Rescue lane** → no link. Reply-CTA only (test ICP presence via replies, not clicks).
+- Hashtags: 2-3 max, relevant, at the end of the body. 0 is acceptable and on-trend. Never a wall of tags.
+
+**Posting the first comment:** claude-in-chrome CAN do this (typing a comment needs no file dialog) - after the post publishes, add the `first_comment` text via the assistant. Only the image attach is a manual gap.
+
+## Analytics tracking — the ledger (proposed "better way")
+
+Instead of scattering metrics across 20 frontmatter blocks, **one reviewable table**:
+[`metrics-ledger.md`](metrics-ledger.md). One row per posted post, filled from
+LinkedIn analytics. This is the weekly review surface (same pattern as the outreach
+`pipeline.md`). Frontmatter only carries `posted_url` + `status`; the ledger owns
+the numbers.
+
+**Review workflow (per post):**
+1. **~48-72h after posting**, open the post's LinkedIn analytics ("View analytics").
+2. Log the row: impressions, reactions, comments, reposts, profile views, and the
+   real signal — **ICP replies** (comments/DMs in ICP symptom-language).
+3. **Weekly (Fri):** scan the ledger. Which opener archetype / lane / topic pulled
+   ICP replies? Reuse the winners; retire the flat ones. Note it in the ledger's
+   "what to reuse" column.
+4. The `icp_replies` count, not impressions, decides whether the campaign is
+   validating (per each lane's plan kill-criteria).
+
+## Posting workflow
+
+1. Draft against the idea-first skeleton; self-score ≤2/10 (pipeline rubric) + the
+   shape-tell critic.
+2. Export the visual to `assets/<slug>.png`. **Attach the image at compose time,
+   BEFORE scheduling** - you cannot add an image to an already-published post
+   (LinkedIn allows text edits only). Scheduling text-first to "add the image
+   later" is how the 2026-08-13 post shipped imageless. (Until claude-in-chrome can
+   drive the native upload dialog, Paul attaches the PNG in the composer.)
+3. Stage the `first_comment` (course lane: UTM'd course link).
+4. **Schedule at least 24 HOURS ahead** (Paul rule 2026-08-13) - never same-day or
+   imminent. The 24h gap is Paul's **pre-verify window**: he reviews the queued post
+   (text, image, shape) before it goes live and can tweak or pull it. Pick the next
+   Tue-Thu US-morning slot that is ≥24h out.
+5. **Paul pre-verifies** within the window. Only then does it publish.
+6. **After it publishes, the assistant posts the `first_comment`** via claude-in-chrome
+   (typing a comment needs no file dialog - this IS automatable). Course link lives
+   here, not the body.
+7. Reply to ICP comments within ~2h; route real conversations to DM.
+8. Log analytics 48-72h later in the ledger.
+
+### Link unfurl (why the course link shows no preview card)
+
+- **Comment links never unfurl on LinkedIn** - only body links generate a preview
+  card. A first-comment link is a bare clickable link by design (the cost of keeping
+  the link out of the reach-throttled body).
+- Separately, our `og:image` is **WebP**. **Telegram unfurls it fine** (confirmed by
+  Paul 2026-08-13) - WebP-friendly crawlers are OK. **LinkedIn is the holdout**: its
+  crawler won't render WebP, so a LinkedIn *body* link shows no preview card. The fix
+  is LinkedIn-targeted: emit a jpg/png og:image in
+  `themes/beaver/layouts/partials/seo/enhanced-meta-tags.html` (or the root override).
+  Not a sitewide emergency (Telegram/most work); worth doing since LinkedIn is a key
+  channel. Tracked as an engineering task.
+
+## Files
+
+- `content-plan.md` — rolling 2-3/wk calendar, lane rotation, revision-wave queue.
+- `reference-examples.md` — transcribed good posts to emulate (add more over time).
+- `metrics-ledger.md` — per-post performance, the weekly review surface.
+- `course-promo/` , `icp-validation/` — the two lanes, each with `README.md`, drafts, `assets/`.