The Agent OS pattern applied to LinkedIn. A workspace rather than a chat window, and 24 skills that compound.
The difference between using a model and having a system is whether anything survives the session. Build in a chat and you forget what you made, cannot find it next week, and rebuild it from scratch. Build in a workspace and every skill stacks on the last one.
Inside: the workspace structure, the P0 to P4 prioritisation that stops everything being urgent, 24 LinkedIn skills written out, the comment-to-DM funnel, and the skillify loop that turns one-off work into something you hand to your team.
Plus where the sending runs, because the one thing a chat window cannot do is reach a person safely.
AI Driven, Personalized Outreach on LinkedIn: Prosp 💜
Setup: A Workspace, Not a Chat Window
The model is not the system. The workspace is.
On its own, a frontier model in a chat window loses everything. You forget what you built, you cannot find it next week, and you spend an hour remaking something that already existed.
Why a workspace rather than a chat
The problem is not capability. It is that nothing survives the session.
What breaks in a chat window:
You build something good on Tuesday.
By Friday it is forty messages up a thread you will not reopen.
Next week you rebuild it, slightly worse, from memory.
What a workspace gives you instead. Every skill, post, sequence and audit sits somewhere you can find it. You stop starting from scratch and start stacking on what you already have.
That one change is the difference between a setup that gets sharper every week and one that resets to zero.
And it is what makes handing work to someone else possible. A skill in a folder can be given to a teammate. A conversation cannot.
The structure
linkedin-os/
├── context/
│ ├── icp.md who you sell to, from closed-won
│ ├── voice.md three of your real posts, in full
│ └── proof.md one row per result, permission column
├── skills/ every skill you have built
├── formats/ the format library, below
├── output/
│ ├── posts/
│ ├── lists/
│ └── sequences/
└── priorities.md P0 to P4, reviewed weekly
Everything that gets built goes in skills/. Not into a chat. That is the whole discipline, and it is the one most people skip because saving feels like admin at the moment you are busy.
P0 to P4
If everything is a priority, nothing gets done. Five levels, reviewed weekly.
P0 Immediate. Blocking revenue this week.
P1 Next actions. Starts when P0 clears.
P2 Queued. Real, not urgent.
P3 Worth doing eventually.
P4 Logged so it stops taking up space in your head.
The weekly review is a skill, not a habit:
---
name: priority-sorter
description: Sort everything in flight into P0 to P4. Weekly.
---
READ the last 7 to 14 days of work in the workspace.
PER ITEM: the project it belongs to, the category, and the
priority level.
OUTPUT sorted P0 first, with a one-line reason per P0.
RULES
No more than three P0s. If there are more, they are not P0s.
Anything at P3 or below for three consecutive weeks moves to P4
or gets deleted. A list nobody acts on is noise.
Flag anything started and abandoned, by name.
NOT YOUR CALL: deciding what the business priorities are. You
are sorting, not setting.
The abandoned-work flag is the useful output. A workspace accumulates half-built things, and naming them is how you either finish or delete them.
The six systems, mapped to LinkedIn
The Agent OS pattern runs six systems. Here is what each becomes when the channel is LinkedIn rather than YouTube and paid.
| The system | On LinkedIn |
|---|---|
| Video | Native video and voice notes in sequences |
| Motion design | Carousels and dense single-image posts |
| SEO content | The post engine: pillars, hooks, drafts |
| Ads | Message variants and angle testing |
| Daily research | Format library and trend scanning |
| Memory | The skills folder, and what it tells you to build next |
The last one is the one that ties the rest together, and it is the least obvious. Looking at what your skills keep doing tells you what to automate next, which is the question most people cannot answer.
If you keep writing carousels by hand, build the carousel skill. If you keep writing the same reply, build the reply skill. The workspace shows you where the time goes.
The skillify loop
The habit that makes the whole thing compound.
1 Do the thing once, properly, with feedback until it is good.
2 Turn it into a skill file while it is still fresh.
3 Put it in skills/ where the team can find it.
4 Hand it to someone who will make it better.
Step 4 is the one that scales. Getting something 70% right and handing it to someone who takes it to 90% is a different operating model from doing it yourself to 95%.
The skill file format:
---
name: skill-name
description: What it does and when to run it. This field is the
whole triggering mechanism, so write it as the sentence
someone would say when they need it.
---
READS: the specific context files. Never the whole folder.
OUTPUT: what it produces, in what format.
RULES: the constraints that make the output usable.
FAIL CLOSED: when to stop rather than produce something
plausible.
Every skill needs the fail-closed line. It is what separates a skill from a prompt, and it is the thing a teammate relies on when they run it without knowing the context you had.
What the workspace cannot do
One thing, and it is the thing that turns all of this into pipeline.
It cannot safely reach a person on LinkedIn. A model driving a browser session on your own account produces activity from an unusual environment, at machine pace, from an IP with no history. That is the fastest route to a restricted account, and the account is the asset.
Prosp is the sending layer: cloud-based so the sequence continues when the laptop is shut, with a free dedicated residential proxy per connected account and activity paced under the platform's thresholds.
The constraint it enforces governs everything in this guide. Roughly 20 connection requests a day and 100 messages a week, account-wide rather than per campaign. Three campaigns set to 20 each means the account attempts 60.
That number is why the skills on the next page are weighted toward content and formats rather than volume. Comments are free. Sends are not.
The 24 LinkedIn Skills
24 skills, four groups of six. Every one saves to skills/ and every one has a fail-closed line.
The format library and the comment funnel are the two worth building first.
Research and formats, 6
---
name: format-library
description: Find the post formats actually working right now and
keep a reference library of them. Updates itself weekly.
---
SCAN for posts getting unusual comment volume relative to
reactions, in your category and adjacent ones.
PER FORMAT capture:
The structural shape, described so it is reusable
Who used it, with the post URL
THE DATE, because format freshness decays fast
The comment-to-reaction ratio
Whether it depends on an image, a carousel, or text alone
OUTPUT a browsable library, sorted by freshness.
RE-RUN WEEKLY and mark anything over 60 days old as stale
rather than deleting it. A format that stopped working is
useful information.
RULES
Rank by COMMENT-TO-REACTION RATIO, not total engagement. High
reactions means agreement. High comments means a fight, and a
fight is where the attention is.
Never copy a post. Capture the structure, not the content.
FAIL CLOSED on a format you cannot describe in one sentence. If
it cannot be described, it cannot be reused.
This is the skill that unblocks content for most people. The blocker is rarely writing. It is staring at a blank page not knowing what shape the thing should be.
---
name: raw-idea-mapper
description: Take a half-formed idea and map it to the formats
that fit it.
---
INPUT: whatever you have. A sentence, a voice note transcript,
a thought.
OUTPUT:
What kind of idea this is: mistake-to-fix, contrarian,
teardown, number, admission, process
The three formats from the library that fit it
A reference post for each, with its URL
The talking prompt: what you would say out loud
The CTA that fits
RULES
Never force an idea into a format that does not fit. Say the
idea is not ready instead.
FAIL CLOSED if the idea is actually three ideas. Split it and
say so.
---
name: creator-analysis
description: Analyse an account that is working, and extract the
workflow rather than the content.
---
INPUT: an account whose audience overlaps yours.
OUTPUT:
Publishing evidence: cadence, formats, what they repeat
The patterns visible across their last 20 posts
What they do that you could do
WHAT THEY DO THAT YOU COULD NOT, and why
Your actual advantage over them, stated honestly
RULES
The point is the workflow, not the topics. Copying topics
produces a worse version of them. Copying a workflow produces
your version of it.
FAIL CLOSED if your advantage field is empty. If there is
nothing you have that they do not, you are not differentiated
and the content will not matter.
---
name: trend-scanner
description: What is the audience arguing about this week.
Monday, before the calendar is set.
---
OUTPUT five tensions, ranked by comment-to-reaction ratio.
Per tension: the question people are arguing about, the
majority position, the minority position, three verbatim
quotes with URLs, and WHAT NOBODY HAS SAID YET.
The last field is the post.
FAIL CLOSED on a tension with no minority position. That is a
consensus, not a debate, and posting into it adds nothing.
---
name: pattern-library
description: Keep the reusable patterns, not just the outputs.
---
AFTER ANY PIECE OF WORK THAT WENT WELL, capture:
What the job was
What made the output good
The reusable shape
What would break it
OUTPUT a pattern file in skills/, not a note in a chat.
THIS IS THE SKILLIFY STEP. Without it the work was a one-off.
---
name: content-gap-finder
description: What your audience asks that nobody in your category
answers properly.
---
SOURCES: comments on your posts, comments on competitor posts,
questions in relevant communities, your own inbound replies.
OUTPUT the five questions asked most often with the thinnest
answers available.
Those five are your next month.
FAIL CLOSED on a gap you cannot answer from real experience.
Writing the thin answer better does not help anyone.
Content production, 6
---
name: dense-image-post
description: The single image carrying a lot of structured
content. Saves and shares well.
---
THE SHAPE: a numbered or gridded list of items, each one or two
words plus a short line, all on one image.
TYPICAL SUBJECTS: the traits of something, the stages of
something, the mistakes in something, the tools for something.
OUTPUT:
The title line
Every item written out in full, never "and so on"
The CTA, as a solid bar at the foot
RULES
Every item must be specific enough to stand alone. "Be
consistent" is filler. "One post a day maximum, because
same-day multiples suppress each other" is an item.
Name the exact colours, the typeface weight, and what bleeds
off which edge. "Less generic" just swaps one default for
another.
FAIL CLOSED if more than two items are filler. The post is not
ready.
---
name: carousel-builder
description: 8 to 12 slides, one idea each.
---
SLIDE 1 carries the whole promise. If someone sees only slide
one, they should know what they get.
SLIDE 2 is the cost of not knowing.
MIDDLE SLIDES: one point each, none needing the one before it.
FINAL SLIDE: one CTA.
RULES
If it does not fit in ten words it is two slides.
Carousels are saved more than shared, and saves compound over
days rather than hours.
FAIL CLOSED on a paragraph on a slide.
---
name: caption-writer
description: Platform-native captions from one piece of content.
---
PER PLATFORM, write to near the character limit rather than
under it, because the limit is a format constraint not a target
to avoid.
LinkedIn: hook in lines 1 to 2, one idea per line, one CTA, no
hashtags in the body.
RULES
Each caption must work alone. Someone seeing only one has not
seen the others.
Never the same caption with the platform name swapped.
FAIL CLOSED if two captions are the same post wearing different
hooks.
---
name: long-to-short
description: One long piece into standalone short pieces.
---
EXTRACT 10 to 12 standalone ideas. One per line. No idea that
needs the others to make sense.
THEN per idea, one post, hook last.
THEN A CROSS-BATCH CRITIQUE: read all of them together and flag
any two making the same point.
That last pass is the one people skip, and repurposing reliably
produces two posts saying the same thing.
FAIL CLOSED under 8 standalone ideas. The source was not
substantial enough to repurpose.
---
name: humanizer
description: The final pass on everything.
---
Numbers not all round. No perfect parallelism in lists. One
genuine break in the flow of thought. Case studies carry one
vague detail beside one specific one. Not every post resolves.
No em dashes.
TEST: could a competent person have written this in five
minutes? If it reads like a committee, that is what it reads
as.
FAIL CLOSED on anything that fails the test. Rewrite rather
than publish.
---
name: claim-checker
description: Every claim traced to a proof row, before publishing.
---
PER CLAIM in the draft: the row in proof.md that supports it,
or FLAG.
RULES
No row, no claim. No proof is a valid answer: write the version
that makes no results claim.
Check the permission column. YES, ANONYMISED or NO.
FAIL CLOSED on any flagged claim. The post does not publish
with it in.
The comment funnel, 6
The play that turns a post into a contactable list. Six skills, run in order.
---
name: magnet-definer
description: Decide what the free resource actually is, before
any post is written.
---
OUTPUT:
What is literally inside, as OBJECTS. Counts, sections,
files.
THE ONE ITEM useful on its own, which the follow-up names.
Who it is for, by exact job title.
THE RELATEDNESS TEST, and it is not optional:
Someone finds this genuinely useful. What do they need next?
If the answer is not what you sell, this is a content piece
with a CTA on it, and the list it builds is not yours to sell
to.
FAIL CLOSED if you cannot list objects. It is not finished.
FAIL CLOSED on the relatedness test. Build something else.
---
name: magnet-post-writer
description: Three variants, three keywords.
---
A ANNOYANCE what is irritating about the current answer
B PAIN what it costs to keep doing it that way
C RESULT the outcome, then the mechanism
Variant C requires real approved proof. If there is none, write
a second angle on A or B rather than inventing one.
THREE KEYWORDS, same semantic family, never numbered.
OUTBOUND / PIPELINE / SEQUENCE, not OUTBOUND / OUTBOUND2.
THE CTA DEPENDS ON CONNECTION STATE:
Mostly connected: Comment "[WORD]" and I'll DM it over.
Mostly not: Comment "[WORD]" and connect so I can send
it. PS. Already connected, just comment.
FAIL CLOSED if the delivery is not planned first.
---
name: capacity-checker
description: Can you actually deliver to everyone who comments?
Run BEFORE posting.
---
INPUT: expected comment volume, connection state of the
audience, daily send budget.
OUTPUT: weeks to deliver to everyone, and the order.
THE MATHS:
100 comments, mostly connected under a week
400 comments, mostly not four weeks
500 comments, mostly not five weeks
RULES
Everyone who asked gets the resource. The scoring decides what
happens afterwards, not whether they get what they asked for.
Never raise the daily limit to catch up.
FAIL CLOSED if the expected volume exceeds four weeks of budget
and no prioritisation is set.
---
name: engager-scorer
description: Decide the delivery order.
---
DEPTH x FIT, each 1 to 3, multiplied. Maximum 9.
DEPTH 1 reaction 2 short comment 3 substantive or repeat
FIT 1 adjacent 2 close 3 exact match
6 to 9 first. 3 to 5 next. 1 to 2 get the resource and nothing
after.
Tag anyone who commented on more than one magnet. Those are the
warmest contacts the play produces.
FAIL CLOSED on an unscored list. At 20 a day the order decides
the quarter.
---
name: delivery-sequence
description: The five nodes that deliver it.
---
1 Import commenters AND likers from the post URL, with future
reactions on so it keeps pulling
2 Condition: is the lead first degree
3 YES: message with the link
4 NO: connection request, then message on acceptance
5 A public reply to every comment
NODE 5 is the one people skip and the one that compounds.
Public replies keep the post alive, which extends the window
the import keeps pulling in.
NO WAIT CARD after the connection request. Acceptance
auto-detects over two weeks.
THE DELIVERY MESSAGE: thanks, the link, one question.
The question turns a delivery into a conversation. Never pitch
here; they asked for a resource.
---
name: magnet-diagnostic
description: 30 days after. Stop at the first failure.
---
1 Low comments the hook, or the resource was not
worth asking for
2 Comments fine,
low delivery the branch or the cap. Mechanical,
not a copy problem.
3 Delivered, no replies the delivery message had no question
4 Replies fine,
no conversations THE RESOURCE IS UNRELATED TO WHAT
YOU SELL. The expensive one, and it
takes three weeks to notice.
5 All fine, nothing
closes price or fit, not the magnet
ALSO: which variant pulled, and whether repeat commenters
converted better. They almost always do.
The delivery half is the part that cannot run by hand at 400 comments. Prosp imports commenters and likers straight from the post URL including reactions that arrive days later, and the five nodes map onto its canvas builder directly.
Audit and handoff, 6
---
name: profile-leakage-audit
description: Where your profile loses people. The LinkedIn
version of a funnel leak audit.
---
FIVE LEAKS, each with the evidence and what it costs:
1 THE HEADLINE truncates around 60 characters on mobile, which
is where most profile visits happen. What survives the cut?
2 THE ABOUT: do the first two lines say what you do, before
the See More?
3 THE FEATURED: is anything pinned, and is it the thing a
buyer needs before a call?
4 THE CTA: is there one, and is it one action?
5 THE PROOF: is there a single number above the fold?
OUTPUT the five, ranked by what each costs you.
RULES
Every connection request drives a profile visit. A weak
headline loses the accept there rather than in the message, and
no amount of sequence tuning fixes it.
MEASURE IT: acceptance for the two weeks before a rewrite and
the two weeks after, same list quality. Nothing else changed,
so the delta is clean.
---
name: funnel-leak-finder
description: Where the traffic your LinkedIn sends actually
leaks.
---
CHECK the destination of every link you post:
How many form fields before they can convert
How far down the page the form starts on mobile
Whether the page promise matches the post promise
How many exits: nav, footer, other offers
Whether the CTA matches the one in the post
OUTPUT the five worst, ranked.
THE ONE PEOPLE MISS: a post promising one thing linking to a
page about another. That mismatch costs more than any copy
change.
---
name: weekly-diagnostic
description: Monday. Stop at the first failure.
---
1 Low reach the engagement work is not being done
2 Reach fine,
low comments the post resolved the tension, or asked
for nothing
3 Comments fine,
low delivery the branch or the cap
4 Delivered, no
conversations the delivery message had no question
5 All fine,
followers flat the profile. Nobody checks this one.
ONE change per week. Change three and a lift tells you nothing
about which one worked.
---
name: handoff-writer
description: Hand a skill to someone else so they can run it
without you.
---
OUTPUT:
What this does, in one sentence
What it reads and what it writes
The three decisions it makes and the rules behind each
WHAT IT CANNOT DO, named
Where it has been wrong before
The last two are what a handoff usually omits and what the
person receiving it actually needs.
FAIL CLOSED if you cannot name a case where it was wrong. You
have not used it enough to hand it over.
---
name: skills-publisher
description: Put a skill somewhere the team can find it.
---
BEFORE PUBLISHING:
Strip anything client-confidential
Replace your proof with a placeholder
Replace your objections with a placeholder
Check the description field reads as the sentence someone
would say when they need it
THE DESCRIPTION IS THE WHOLE TRIGGERING MECHANISM. A skill
nobody finds is a skill nobody runs.
---
name: what-to-build-next
description: Read the workspace and say what to turn into a
skill. Monthly.
---
SCAN the last month of work.
OUTPUT:
What you did more than three times by hand
What took longest relative to its value
What you rebuilt because you could not find the first version
The three things worth skillifying, ranked
THIS IS THE SKILL THAT ANSWERS "where do I start". Most people
cannot automate the next thing because they do not know what
the next thing is. The workspace knows.
That last skill is the one the whole pattern exists for. A chat window cannot tell you where your time went. A workspace can.
The weekly loop
MON priority-sorter. Three P0s maximum.
MON trend-scanner and format-library refresh.
MON weekly-diagnostic. One change.
TUE raw-idea-mapper on whatever you have.
TUE Draft, humanise, claim-check. Publish.
WED The magnet post, if it is a magnet week.
WED capacity-checker BEFORE posting it.
THU engager-scorer and delivery on last week's post.
THU Second post.
FRI pattern-library on anything that went well.
FRI Third post if the week earned it.
MONTHLY what-to-build-next. Then build one of them.
One magnet a week, not every post. A magnet on every post trains the audience to wait for the free thing and ignore everything else you write.
Three things to remember
- Save everything into the workspace, not into a chat. The compounding is the whole point, and it only happens if the work survives the session.
- Run the relatedness test before building a magnet. A resource unrelated to what you sell produces a large list you cannot convert, and it looks like your best post for about three weeks.
- Get something to 70% and hand it over. Doing it yourself to 95% is a worse operating model than having two people take it to 90%.
The workspace builds it. Prosp reaches people 💜
