The skills

What a skill is, and why it matters to you

A skill is a page of rules that several commands share. Platform limits live in one file; how often to post lives in another; how the plugin talks to you lives in a third. Commands read them, so /plgn month and /plgn post cannot end up following different rules about the same thing.

This matters to you for two reasons. The behaviour you see is not improvised, so it is predictable. And when the plugin does something you disagree with. There is a specific rule behind it, which makes the disagreement a concrete one rather than a complaint about an assistant.

How it talks to you

Reply style

The rules for everything printed in your terminal: reports, plans, questions, errors and progress lines. This is about the plugin's own voice, never about your brand's: the two are kept strictly apart, and a brand's voice is never applied to a reply.

• Reply in the language you wrote in. Arabic in, Arabic out. Command names, brand names and platform names stay unchanged.
• Keep it simple in every language: sentences of about fifteen to twenty words, one idea each, the common word rather than the rare one, and active voice.
• Sound like a good colleague. Not "Operation completed successfully. 28 entities created", and not "Amazing! Your month is all set!", just "Your month is scheduled, 28 posts, 24 with images. Two need a look."
• Make the reply as long as the job. One post created means one line back.
• Result first, question last. Never explain what it is about to do; do it, then say what happened.
• Say what was assumed, in one line, because you cannot correct an assumption you cannot see.

The words you should never see

A list of internal vocabulary that must never reach you, with the plain replacement for each. "The validation gate refuses it" becomes "plgn's checks blocked it". "A starved pillar" becomes "a topic with too few posts". Tool names, worker names and error codes are simply never printed.

Three internal words are deliberately allowed: Foundation, Business and Creative. Your dashboard prints those three as headings, so a reply that avoided them would be describing a screen you cannot find.

The closing line on free commands

Free commands end with one short block naming what connecting would add: and there are exactly two versions of it, because using the wrong one produces nonsense. A command that drafted posts says those posts are sitting still. A command that produced findings says the findings are not acted on. An audit has no posts to schedule.

It appears once per command and once per session. Three near-identical pitches in a row is noise, and it never appears in a connected command at all.

How it writes for you

Brand voice

How to write as your brand: match the rhythm and the words, not just the subject.

• Sentence length and variety. Copy the pattern of long-then-short, not just the average.
• Their words for their things. If you say "workspace". It never says "account".
• What the brand never does, often more telling than what it does. A brand that never uses exclamation marks or never opens with a question has a voice built on those refusals.
• Banned words are absolute, and a draft that respects them from the start is a draft nobody has to fix.
• A voice is language-specific. A brand that sounds plain and direct in English does not automatically sound that way when the same rules are applied to Arabic. The intent is matched in each language rather than the rules translated.

When no voice is saved, which is every free command, it works one out from your homepage, about page and one product page, and says so in one line before showing you anything. A worked-out voice is a guess from a few pages, and you should know which parts of the output rest on it.

Platform limits and habits

The target matters more than the limit. A LinkedIn post at 2,995 of 3,000 characters fails the moment anyone edits a word, so posts are written to the target and the check has room.

Content topics

A topic is a repeated argument, not a category. "Educational" is a category and tells a writer nothing. "Migration stories" is an argument: old tools trap teams, and we have the receipts.

• The test: could this topic belong to a competitor unchanged? If yes, it is not a topic.
• Three to five. Fewer and the feed repeats itself within a month; more and no topic gets enough posts to build a pattern.
• Every post must fit exactly one topic, with no ties. If two topics would both claim the same post, they are one topic written twice.
• Across a month, roughly half should help the reader whether or not they buy, a third should be proof, and a fifth direct asks. These are proportions, not quotas: but a feed permanently flipped, where every post asks for something, is the one that burns an audience.
• The strongest topic is usually an argument your competitors all avoid. Where everyone says the same three things, the fourth is free.

How often to post

• No topic takes a whole week. Posts rotate so one follows a different topic.
• Each platform is spread evenly: three LinkedIn posts are Monday, Wednesday, Friday, not three in one afternoon.
• The last week is left lighter, because something always takes it. A month with no slack is a month that breaks.
• Times are sensible defaults, never promises. Real best times can only be known from your own data, which plgn does not hold, so it never claims a time gets more attention.

How it learns your brand

Where a brand's knowledge lives

The map of the four places, the three layers and the sixteen kinds: which has its own page here, because it is the thing most worth understanding. The short version: banned words go on the brand record, what you sell is an offering, and putting either in the wrong place fails quietly.

Capturing a brand

The method /plgn setup and /plgn brandkit follow. One rule governs the whole thing.

• Sources are used best-first: the brand's own published posts, then the website, then any style guide, then competitors: and never a competitor for voice, because borrowing it makes the brand sound like the market.
• Every claim carries its evidence and a confidence. No evidence, no save: a claim with nothing behind it becomes a question, never a saved entry.
• At most five questions, each arriving with an answer already drafted. A beginner says yes to five answered questions; an expert edits four and finishes just as fast. Neither is handed a blank page.
• Anything past the fifth is printed as a stated assumption rather than guessed in silence.
• The bar an entry has to clear: could two writers follow this and produce opposite copy? "Professional and friendly" fails. "Short sentences, no exclamation marks, never opens with a question" passes.

How a brand looks

Working out a visual direction from real pictures, storing it, and applying it before an image is made. The ten parts of a direction are on the page about what plgn knows. The rules around it are here:

• Three references minimum. Fewer is a sample, and it says so rather than dressing a guess up as a direction.
• A picture nobody could see is not evidence. It never describes an image from its filename or from the caption of the post it belongs to.
• When references disagree, it reports the groups and asks, averaging two identities produces a third that belongs to nobody, and every image afterwards is wrong in a way nobody can name.
• Honest limits are stated: colours read off a compressed screenshot are close, not exact, so it says "about this" rather than publishing a value as though it came from a brand book.

The brand's own things

The real things a brand owns, and how a picture is built around them instead of inventing them: the logo, a mascot, the founder, the shop, a pattern. A description of a mascot makes a new mascot in every picture. The mascot's own picture makes the mascot. This is the rulebook behind /plgn assets.

• Each thing is saved with its pictures, what it is, when to use it, and a list of what it must never do. A logo is never stretched. A mascot is never angry. That list is the part people forget.
• A real face needs consent. It is used in a made picture only when you have said the person agreed. It is never assumed, not for the founder and not for you.
• Logos, templates and badges are not given to the image model at first, because it redraws letters wrong. The owner can turn that on for one thing after checking the results.
• Product photos are not saved here. They belong to the product, so there is only one copy.

How it makes pictures

Three separate things: deciding what the picture is about, writing a description worth spending points on, and handling the wait that follows.

Deciding what the picture is about

Before any picture is made, the thinking behind it is written down. This record is called a brief. It is what /plgn why reads back months later, and it stops the same idea being offered to the same brand twice.

1. The benefit: take the one thing the post promises and say what it could really mean to someone.
2. The ideas: turn those meanings into a few ideas for the picture and pick one. Every losing idea is kept, with the reason it lost.
3. The direction: say what each frame shows. What carries the frame, a real photo, a built object, a scene or type alone, comes from what the brand sells, not from taste.
4. The check: the designer checks each frame against the brand's rules, then writes the final image description.

Writing the description

1. Subject: what is in the picture, concretely. "A desk with one open notebook and a closed laptop", not "productivity".
2. Framing: the viewpoint, where the subject sits, how much empty space. Roughly square survives more crops than wide.
3. Light: direction, hard or soft, time of day. This sets the mood more than any adjective.
4. Mood: tied to what the post argues, not what it is about.

It also avoids drawing the caption. A post about wasted planning time does not need a picture of a calendar: the literal illustration is the first idea and almost always the weakest, because it adds nothing the reader just read. Empty chairs after a meeting says more about wasted meetings than a clock does.

When to make nothing

• The post is a written argument that reads stronger plain, common for long LinkedIn posts.
• The only description possible is generic, and the result would look like stock. A stock-looking image costs points and some credibility.
• There is no image setup yet. It says so and does not retry.

Spending no points is a valid outcome, reported as a decision rather than a failure.

The wait

• Asking for a picture returns a job, not an image. It is checked every five seconds.
• It stops checking after about ninety seconds per picture. A picture still waiting in the queue is reported as still running, not as failed, so you can check it later. That post waits without one for now, and the run carries on. A missing picture never blocks scheduling.
• When several are needed, all are started first and then checked, one at a time turns a ninety-second worst case into half an hour.
• It never remakes a picture just because the first was dull. That is paying the points twice for a small gain. A remake happens only when one genuinely failed.
• Every picture gets alt text: one sentence describing what is visibly there, starting with the subject, never repeating the post.

How it recovers

The rules for what happens when the server refuses something. The behaviour is on the page about how it works; what belongs here is what does not count as recovering.

Three other refusals are not about your content at all and need different handling: a permission your role does not have, an item that has moved since earlier in the run, and something not connected yet. None of them is fixed by rewriting copy, so none of them causes a rewrite.