agentleFS
thescottkingthescottking/b2b-pmm-skills/skills/social-generator/SKILL.md

social-generator

Skill0 starsChanged 26 days agoMITView on GitHub

Write social captions and matching image prompts from an article, a URL, or a topic brief. Use whenever someone asks for a LinkedIn post, an X post, an Instagram caption, a text overlay for a graphic, or an AI image prompt for a social graphic. Triggers on "write a LinkedIn post about this", "turn this article into social", "make an Instagram caption", "generate a Midjourney prompt for this post", "I need a week of social from this blog", "give me a text overlay for the graphic", or any request pairing a piece of content with a platform name.

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. About 0 are common, 3 are distinctive, and 10 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 10

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read these first. They are the source of truth and this skill does not invent around them.

    line 23
  2. Very specific:

    If brand-kit/ is missing or still has its template placeholders, stop and tell the user to run brand-kit-setup.

    line 36
  3. Very specific:

    If house-style is installed, run it as the final voice pass over every caption before delivery.

    line 39
  4. Very specific:

    Hashtags come from Hashtags in brand-kit/brand.md: Never invent one. That section splits them into branded and topical and gives a …

    line 53
  5. Distinctive:

    No text overlay needed.

    line 115
  6. Very specific:

    Read Illustration Style in brand-kit/brand.md and build both prompts from it.

    line 150
  7. Very specific:

    If the brand kit has no Illustration Style section, stop and help the user write one using the guide below, then add it to …

    line 153
  8. Very specific:

    Every value here is a role from Palette in brand-kit/brand.md.

    line 201
  9. Distinctive:

    Check contrast before you commit.

    line 219
  10. Very specific:

    Add exclusions that contradict the style: gradients and color fills for a line style, outlines for a flat-shape style, photography for both.

    line 249
  11. Very specific:

    Every claim traces to What We Do in brand-kit/capabilities.md.

    line 309
  12. Very specific:

    Check Superseded Names in brand-kit/voice.md before writing any product name into a caption.

    line 312
  13. Very specific:

    If brand-visuals is installed, mark the accent words in the graphic headline with asterisks, *like this*, and offer to render the headline …

    line 314

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Social Generator

Produce a complete social package from one source: platform-native captions, the bold text overlay that goes on the graphic, the graphic headline, and paired image prompts for Midjourney and DALL-E.


Before you write anything

Read these first. They are the source of truth and this skill does not invent around them.

  1. brand-kit/brand.md: name, palette, fonts, CTAs
  2. brand-kit/voice.md: house style and banned constructions
  3. brand-kit/positioning.md: the frame, pillars, traps, stats, vocabulary
  4. brand-kit/capabilities.md: read before making any product claim

Two shared references carry structure this skill does not own:

  1. docs/writing-templates.md: the LinkedIn post beats and its length target
  2. docs/messaging-patterns.md: the eight structural patterns, one per caption

If brand-kit/ is missing or still has its template placeholders, stop and tell the user to run brand-kit-setup. Do not proceed with invented positioning.

If house-style is installed, run it as the final voice pass over every caption before delivery. Do not hand-copy voice rules into this skill.


Two rules that get broken most often

Never name the framework. The qualities in The Pillars in brand-kit/positioning.md are internal scaffolding. Pick one to lean on and let it shape which claim you lead with, then throw the label away. A caption that says "our three pillars", "our messaging framework", or any pillar name by title has failed. AI Tells to Avoid in brand-kit/voice.md calls this scaffolding exposure and the quality panel penalizes it.

Hashtags come from Hashtags in brand-kit/brand.md. Never invent one. That section splits them into branded and topical and gives a count per platform; use those tags and those counts. If the section is missing or its table is empty, ask the user for the set once, tell them it belongs under Hashtags in brand-kit/brand.md so every skill finds it, and ship the posts without hashtags in the meantime. Do not guess at a branded tag that may not exist. If the section says in words that the company does not use hashtags, that is an answer: ship without them and do not ask again.


Step 1: Gather inputs

Seven questions. Ask for anything the user has not already given you. If they gave you a URL and nothing else, ask the rest in one message rather than one at a time.

  1. Source. An article URL, a pasted draft, or a topic brief. If it is a URL, read it and pull the title, the core argument, the target reader, and every concrete number.
  2. Platforms. LinkedIn, X, Instagram, or any combination.
  3. Angle. Which quality from The Pillars or which entry from Named Traps in brand-kit/positioning.md this post leans on, and which pattern from docs/messaging-patterns.md delivers it. One pattern per caption. Internal answer only.
  4. Tone. Authoritative, pragmatic, contrarian, or plain. Stays inside Tone in brand-kit/voice.md either way.
  5. Background. Which palette role grounds the image. See the table in Step 3.
  6. Subject. What the illustration depicts. Name the actual objects, not a category.
  7. CTA. Lift it from Calls to Action in brand-kit/brand.md, social context row. Optional on X and Instagram, expected on LinkedIn.

Surface the numbers you found from the source and confirm which one leads. A real statistic outperforms a general claim on every platform. Check any number you plan to state as a house claim against Headline Stats in brand-kit/positioning.md. If it is not there, attribute it to the source instead of asserting it.


Step 2: Write the captions

LinkedIn

  • Length and beats come from LinkedIn post in docs/writing-templates.md. Do not restate them here and do not invent a character budget.
  • Lead with a real statistic from the source. Not a question, not a setup line.
  • One idea. A LinkedIn post that makes two arguments makes neither.
  • Name the company once, naturally, and never as the subject of the sentence.
  • Article link at the end. The template allows an empty CTA beat; this skill does not, because a social package ships a destination.
  • 3 to 4 hashtags from Hashtags in brand-kit/brand.md, one branded.
  • Text overlay: a bold 3 to 7 word phrase for the graphic.

X

  • Under 280 characters, counted. State the count.
  • One sharp claim or one number. Nothing survives compression except the point.
  • 1 to 2 hashtags from Hashtags in brand-kit/brand.md, topical unless the post is company news.
  • Link optional and it costs characters, so earn it.
  • No text overlay needed.

Instagram

  • Hook on the first line. It is the only line most readers see before the fold.
  • Around 400 characters total.
  • 2 to 3 sentences of substance after the hook, then the CTA.
  • 6 to 8 hashtags from Hashtags in brand-kit/brand.md, mixing branded and topical in the proportion that section sets.
  • Text overlay: a bold 3 to 7 word phrase for the graphic.

Text overlay and graphic headline

These are two different things and both get delivered.

  • Text overlay is 3 to 7 words. It sits on the image. It is the claim, not the topic. It reads at thumbnail size on a phone.
  • Graphic headline is 4 to 8 words. It is the punchier editorial line that can run above or below the art in a carousel or a link card. Mark which one or two words take the accent treatment, using the Accent role from Palette in brand-kit/brand.md.

Both obey Banned Words and Phrases in brand-kit/voice.md. Short copy is where banned words hide best, so scan them specifically.


Step 3: Write the image prompts

Produce a Midjourney prompt and a DALL-E prompt for every graphic. They are not interchangeable. Midjourney takes comma-separated fragments and flags. DALL-E takes plain declarative sentences and ignores flags entirely.

Read the illustration style from the brand kit

Read Illustration Style in brand-kit/brand.md and build both prompts from it. That description, not this skill, defines what the picture looks like.

If the brand kit has no Illustration Style section, stop and help the user write one using the guide below, then add it to brand-kit/brand.md under its own heading. Do not substitute a house style of your own invention. Two different graphics generated from two different guesses is worse than waiting.

How to write an Illustration Style

Four decisions. Write two or three sentences on each and the description is done.

  1. Line weight. Thin and precise, medium and even, or heavy and confident. Is there hand-drawn wobble or is every line mechanically straight.
  2. Fill. Outlines only, flat single-color fills, two-tone blocking, or full tonal rendering. This single choice does more to define a visual signature than any other.
  3. Subject treatment. What the picture shows and how literal it is. Diagrammatic icons, isometric objects, abstract shapes, human figures, photographic composition. Say what is never depicted too.
  4. Edge behavior. How the art meets the frame. Bleeds off one edge, contained with a margin, vignetted, or full-bleed edge to edge.

Three contrasting worked examples. Any of these is a valid answer, and the point is that they are visibly different from each other.

Example A: technical line diagram. Thin, precise pen lines with slight hand-drawn imperfection, roughly uniform weight throughout. Outlines only, no fills, no gradients, no shading. Subject is an interconnected system diagram in isometric perspective: nodes, arrows, gears, documents, servers, queues. Never people, never photographic objects. The drawing bleeds off the lower-right edge with no border or frame.

Example B: flat editorial shapes. No outlines at all. Form comes from flat, hard-edged color blocks in two palette colors plus the background, with no line work anywhere. Subject is a single oversized metaphorical object per graphic, cropped tight, rendered as simplified geometry. Human figures appear as silhouettes only, never with facial features. The composition sits inside a generous even margin on all four sides.

Example C: rendered dimensional objects. No visible outlines. Soft-shadowed three-dimensional forms with matte surfaces and a single consistent light source from the upper left. Subject is one abstract sculptural object per graphic: stacked planes, tubes, spheres, nothing that names a real product. No text inside the art. The object floats centered with clear space around it, full-bleed background behind it.

Background, line color, and text color

Pick a background role, then take the line and text colors from the same row. Every value here is a role from Palette in brand-kit/brand.md. Look up the hex there at generation time. Never hardcode a hex into this skill or into a saved prompt.

Background roleLine colorsText colors
BackgroundBody text primary, Accent secondaryBody text headline, Accent highlight
Background softBody text primary, Primary secondaryBody text headline, Primary highlight
Primary darkAccent primary, Accent soft secondaryBackground headline, Accent highlight
PrimaryBackground primary, Accent soft secondaryBackground headline, Accent soft highlight
AccentPrimary dark primary, Background secondaryPrimary dark headline, Background body
Accent softPrimary dark primary, Primary secondaryPrimary dark headline, Primary highlight
Muted text as a mid-tone groundBackground primary, Accent secondaryBackground headline, Accent highlight
Success / positivePrimary dark primary, Background secondaryPrimary dark headline, Background body

Two rules on top of the table:

  • Two line colors maximum. A third reads as decoration.
  • Check contrast before you commit. If the text color and the background role resolve to similar values in the brand kit, pick a different row rather than shipping an unreadable overlay.

Midjourney prompt structure

Fill the brackets from the brand kit and the user's answers.

[background role color], [illustration style summary from the brand kit],
[line color roles resolved to their hex], [fill rule from the style],
[subject: name the actual objects], [edge behavior from the style],
[composition notes] --ar 1:1 --style raw --stylize 450 --no photography, gradients, text, watermark, logo

What the flags do. Get these right or the output drifts.

  • --ar sets aspect ratio. --ar 1:1 for a feed square, --ar 4:5 for the taller LinkedIn and Instagram portrait crop, --ar 16:9 for a link card or a video thumbnail. Set it per platform, not once.
  • --style raw turns off Midjourney's default aesthetic pass. Use it for any graphic that has to match an existing brand look. Omit it only when you want the model's own taste, which for brand work is almost never.
  • --stylize controls how far the model may depart from the prompt, from 0 to 1000. Low values follow the words literally and look plainer. High values look better and obey less. Around 400 to 500 keeps a defined illustration style intact while still producing a usable image. Move it down when the model keeps adding elements you did not ask for.
  • --no lists what to exclude. Always exclude text and watermark, because generated lettering is unusable and you are adding real type in the overlay. Add exclusions that contradict the style: gradients and color fills for a line style, outlines for a flat-shape style, photography for both.

These flags use a double hyphen. That is Midjourney syntax, not punctuation, and it is the one place the no-em-dash scan should not flag a double hyphen.

DALL-E prompt structure

Plain sentences. No flags, no comma-fragment stacking, no weights. State the constraints positively where you can, because negations are followed loosely.

[Background role color] background. [Illustration style described in full
sentences from the brand kit.] [Line colors named by their resolved hex and
what each is used for.] [Fill rule stated as a positive instruction.]
[Subject: the actual objects and their arrangement.] [Edge behavior.]
[Two or three exclusions at the end, phrased simply.]

Set the size in the request rather than the prompt: square for a feed post, portrait for LinkedIn and Instagram, wide for a link card.


Output Format

Group by platform, then the image section.

## LinkedIn
**Text overlay:** [3 to 7 words]
**Caption:** [beats and length per `LinkedIn post` in `docs/writing-templates.md`]
**Hashtags:** [3 to 4 from `Hashtags` in `brand-kit/brand.md`]

## X
**Caption:** [under 280 chars] (count: __ / 280)
**Hashtags:** [1 to 2 from `Hashtags` in `brand-kit/brand.md`]

## Instagram
**Text overlay:** [3 to 7 words]
**Caption:** [~400 chars, hook first]
**Hashtags:** [6 to 8 from `Hashtags` in `brand-kit/brand.md`]

## Image
**Graphic headline:** [4 to 8 words, accent words marked]
**Palette row:** [background role, line roles, text roles, with resolved hex]
**Midjourney:** [prompt with flags]
**DALL-E:** [prompt in sentences]

Notes

  • A real number from the source is the strongest opening line on every platform. Confirm which one leads before writing anything.
  • Name specific icons in the image prompt. "Systems" produces mush. "Ticket queue, routing arrows, headset, timer" produces a picture.
  • Regenerate the palette row per graphic. A campaign that reuses one background for six posts reads as one post seen six times.
  • Every claim traces to What We Do in brand-kit/capabilities.md. Short copy is where unsupported claims slip through, because there is no room for the qualifier that would have made them true.
  • Check Superseded Names in brand-kit/voice.md before writing any product name into a caption.
  • If brand-visuals is installed, mark the accent words in the graphic headline with asterisks, *like this*, and offer to render the headline and overlay as finished square or portrait cards. It reads the same Palette roles as the table in Step 3, so the card and the image prompt stay on one palette row. Image prompts still go to whatever image model the user runs.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.