LinkedIn Publisher

Draft a post from project facts, then publish straight to my own LinkedIn feed — without pretending I have analytics I cannot reach.

What it is

A LinkedIn integration built into log.exe. It does two things: it turns a project row in D1 into a paste-ready caption, and it publishes that caption to my own feed.

The generation side was the easy half. The honest half was harder.

The OAuth part

LinkedIn's developer app gives you a narrow, well-defined set of grants. I registered one and wired the standard 3-legged flow:

• openid, profile, email for identity
• w_member_social for posting

The OpenID sub is what matters — it becomes urn:li:person:{sub}, the author URN the UGC Posts API wants. Publishing goes out with X-Restli-Protocol-Version: 2.0.0, and a successful post stamps lastPostedAt, which is the single fact the comeback coach computes its whole gap analysis from.

Tokens last 60 days. This app has no programmatic refresh, so rather than let it break quietly two months from now, the UI shows the expiry date and tells me to expect a reconnect.

What the API will not give me

This is the part I want to keep visible.

I cannot read my own analytics. Profile views and post impressions need r_member_profileAnalytics and r_member_postAnalytics, which ride on the Community Management API. On this app, the Request access button is greyed out — it is not a matter of waiting for approval, I cannot even apply.

So every number in this feature is one I pasted in by hand from LinkedIn's own UI, and the page says exactly that where the numbers appear. The suggested posting windows are labelled Starting guess (MYT) … Not from your analytics, because they are a fixed rule of thumb and nothing else.

It would have been easy to render those windows next to a "connected" badge and let them read as insight. They are not insight.

Facts, not fluency

Captions are assembled from what the CMS already knows: title, tagline, description, highlights, tech stack, attached media. If the project has ten highlights, the post can draw on them; if it has none, the template says less rather than inventing more.

Two rules earn their place:

• A private repo's link never ships. quack-quack is private, so a github.com/…/quack-quack line 404s for every reader. The template used to re-emit it on every regenerate, silently undoing a fix I had made by hand. Now the link only appears when the project is public.
• The model cannot write the clock. NIM may reword tone, plans and hook ideas. It may not produce posting times — a rewritten Wed–Fri 15:00–20:00 is a number with no source, which is precisely what that same block warns against.

Voice as data

I scraped my own LinkedIn posts and wrote down what I actually sound like: simple English, first-person, build logs rather than product blurbs, Melayu reserved for community posts, 150–280 words, 3–4 niche hashtags.

That lives in docs/linkedin-voice.md and is compiled into the template — not appended to a prompt and hoped for. The template hits the shape reliably. The optional NIM pass is what gets closer to the feel, and a badge always tells me which one I am reading.