How it works

The four layers

commands/   what you type     : one job each, owns the question it asks you
     |
agents/     parallel workers  : fresh memory each, hand back findings only
     |
skills/     shared rulebooks  : imported by many commands, one source of truth
     |
the server  plgn itself       : checks, storage, and who is allowed to see what

The loop everything follows

1. 1ReadThe command checks the connection, then reads your brand in one go: the record, the voice and audience, what you sell, and any campaign running now, always in that order.
2. 2PlanIt works out what it is about to do and shows you: which topics, how many posts, how many pictures, what they cost, which dates. Then it stops.
3. 3AskOne clear question, one clear answer. Nothing has been written yet. --dry-run ends the run here.
4. 4WriteWorkers run at the same time, one per topic. Each is given everything it needs inside its own instructions.
5. 5CheckEvery post is checked by the server as it is saved. Anything that fails comes back with a reason.
6. 6ReportCounts first, then the exceptions by name. You are told what needs your eye and what does not.

Every command that writes follows this shape. It is why running an unfamiliar command is safe: you always see a plan before anything changes, and --dry-run always stops at step three.

The checks, and why they are on the server

Character limits and your banned words are checked by plgn's server, not by the assistant. A post cannot be scheduled or published while a check fails, whatever the assistant thinks of it.

What happens when a check refuses a post

1. 1It reads the reasonThe refusal names what failed: too long, or a banned word.
2. 2It fixes that one thingToo long means tightening the writing, cutting qualifiers and repeated points first. Never cutting a sentence off mid-way, and never dropping the closing ask, which is the reason the post exists.
3. 3It tries once moreOnce. Not three times.
4. 4Then it hands it to youIf it fails again the post stays a draft and appears in the report by name, with the reason in a normal sentence.

One retry, not five, for a specific reason: being too long is arithmetic and one honest pass fixes it, but a banned word is editorial. If the first replacement does not work, the post's point disagrees with what the brand stands for, and that is a decision for a person. Repeated attempts also make writing worse, each pass strips out more voice to satisfy a rule.

Refusals that are instructions

Not every refusal is a problem. Five of them are the server telling the plugin what to do instead, and you will usually never see them at all.

Why a month takes minutes

A month is not written one post at a time. One writer is started per topic, all at once, so three topics are written in roughly the time one would take.

Pictures work the same way. All of them are started first, then checked on. Making them one at a time, make, wait, make, wait, turns a ninety-second worst case into half an hour.

What a worker can and cannot see

Each worker starts with a clean memory. It cannot see your conversation, the rulebooks, or what the other workers are doing. Everything it needs, the voice, the banned words, the character limits, the campaign's message, is written into its own instructions by the command that started it.

• This is why the plugin passes rules around rather than pointing at them: a worker told to "follow the platform rules" would have nothing to read.
• Workers never save anything. They hand back what they wrote, and the command does every write. That keeps one place responsible for what reaches your workspace.
• A worker that finds a topic thin returns fewer posts and says so, rather than padding to hit a number.

How your brand is read

Any command that writes in your voice makes one read of your brand rather than assembling it from pieces. That read always comes back in the same order, and the order is the priority.

1  the brand record   name, languages, banned words, timezone
2  Foundation         who you are, how you sound, who you talk to, how you look
3  Business           what you sell, your proof, objections, competitors
4  Creative           references and posts worth imitating

The read is also shaped by who is asking. A copywriter is given the voice and the benefits; an art director is given the colours and the picture rules. Each gets what it needs and not the rest, which keeps a writer from drowning in a palette and an art director from being handed a tone guide.

One more thing comes from that read and is easy to miss: your languages. They belong to the brand, not to the conversation. Someone who writes to plgn in English may publish only in Arabic, and the posts follow the brand.

The rules every command follows

1. Reply in the language you wrote in. Arabic in, Arabic out. Command names, brand names and platform names stay as they are.
2. Check the connection before doing anything, and say nothing when it passes. A check that succeeds is silent.
3. Show the plan and wait for a clear yes before creating, changing, scheduling or removing anything.
4. Confirm anything that removes something by name, never by number.
5. State the point cost inside the question, before spending.
6. Say what was assumed, in one line, so you can correct it.
7. Never ask you to type, paste or store a key, a token or a password.
8. Never print an internal name, an error code or a tool name at you.
9. Lead with the thing you asked for, and put the process behind it.
10. Never stop halfway in silence. If a run breaks, say exactly what is in the workspace now.

What it will not do