Lesson 6 of 10 · 8 min
Use AI to draft, and a person to post
Writing five specific comments a day, every day, is where most people quit. An AI model is good at the part that makes it slow: reading a long post carefully and producing a first draft that engages with what it actually says. It is bad at the part that makes it work: knowing what is true about your business, and knowing when a draft sounds like a machine.
The prompt
Paste your one-page offer document from lesson two. Then the rules from lesson four, in full: two short paragraphs of specific value, a free ask tied to this post, under 120 words, no link, no compliments, no “great post.” Then the post itself, copied from the page. Ask for three drafts, not one, so you are choosing rather than accepting.
The batch route
The video does not draft one at a time; it has Claude Code read every post in the sheet from lesson three and write the comment beside it, five hundred in one go. The prompt is the same as above with one change: instead of one pasted post, point it at the sheet and ask for a draft in a new column per row. What makes his drafts work, and he says this twice, is that the model is given a lot about the offer: who it is for, what the group or the sheet contains, what people get out of it. That is why every ask on his screen is different: “we compare LinkedIn format tests like these against reply rates,” “SDRs compare qualification questions against booked-to-held numbers.” A thin offer document gets you five hundred copies of the same comment, and that is the one thing the spam filter is certain to notice.
A ready-made brief: linkedin-agent-skill
If you would rather not write the prompt yourself, there is a free set of Claude Code skills that already contains a better one. linkedin-agent-skill (MIT, eleven skills) has a /li-comment command that takes a pasted post and returns two comments of different types, labelled, with a line on which to post. It works from nine comment types and picks by what the post is: add a number the post lacks, add the missing case, disagree after agreeing first, extend one line, ask the real question, give the receipt from having done the thing, correct a fact, reframe, or a one-liner under twelve words. Its rules are lesson four’s, stated harder: two to four sentences, never open with “Great post” or the author’s first name and an exclamation mark, never restate the post, one idea. It also has a batch mode that takes five to ten pasted posts at once, which is how to feed it rows from the lesson three sheet.
The other skill worth having is /li-human, two Python scripts that score a draft on five checks (sentence-length variation, specificity, stock phrases, invisible characters and em dashes, voice) and strip the machine tells. Install: clone the repository and copy skills/li-* into ~/.claude/skills/, or run /plugin marketplace add Jakeschincariol/linkedin-agent-skill then /plugin install linkedin-agent. Then fill in voice.md, ten minutes, or every user of the skill sounds like every other user of the skill.
The rest of the kit, and where each piece fits
/li-profile: the rubric in lesson five./li-commentand/li-human: this lesson./li-dmand/li-reply: the connection note and your own threads, lesson eight./li-auditand/li-plan: reading what worked and planning the week, lesson nine./li-post,/li-carousel,/li-repurposeand/li-inboxare for your own posting, which this course does not need: commenting works with no audience of your own.
Checked We installed it and ran the humaniser on two drafts before recommending it. The comment brief is good; the cleaner needs reading after.
The test is written up on the blog: we ran the LinkedIn agent skill’s humaniser on our own posts. On a real post of ours it scored 66 and flagged sentence rhythm, which a human editor would have said too. On a deliberately bad draft it removed ten stock phrases and an em dash, and left “Right now of e-commerce” behind where “in today’s fast-paced world of e-commerce” had been, because a regex cannot know what a phrase was attached to. Read every output before you paste it. The skill will not scrape and will not post; it says so in its own README, for the same reason lesson three tells you to be careful. It drafts. You post.
What you do with the draft
- Check every fact. The model will invent a plausible number if you let it. Replace it with yours, or cut it.
- Cut the first sentence if it restates the post. It usually does.
- Rewrite the ask in your own words. It is the sentence people reply to; it has to sound like you.
- Post it yourself, from your own account, by hand.
Checked The batch drafting is the source’s, kept. Posting by hand is the one line we hold, because it is the step that gets accounts flagged.
In the video Claude Code reads each scraped post and writes the comment, five hundred at a time, into a sheet. He never shows anything posting for him, and we would not add it: LinkedIn’s User Agreement 8.2 item 13 names bots that “comment on” posts, and its Community Policies name “gratuitously repetitive” content. A sheet of drafts is neither. A script that posts them is the first, and five hundred near-identical asks is the second, and both are what a restriction looks like from the other side. Draft in bulk if you like; post one at a time, yourself.
Spotted something wrong in this lesson? Tell us, and we will check it and fix it.