# Reference Style Model

> "We want our LinkedIn to sound like them." Turns a reference company's public posts into a style spec: numbered moves, each proved by a real line, and a pre-publish checklist. It transfers form, never their facts.

Reference Style Model is an AI skill for sales & marketing and creative & design teams that you can use in Claude Code, Codex, or type.com. It works with LinkedIn.

- Skill page: https://type.com/library/skills/reference-style-model
- Raw SKILL.md: https://type.com/library/skills/reference-style-model/SKILL.md
- Created by Barbara Jovanovic from Startup Cookie
- Updated: October 30, 2026
- Package: 1 file

## What it does

- Converts a vague brief into 8–12 moves and a yes/no checklist that every draft runs through.
- Gap analysis shows what you already do and what's genuinely new: usually a small shift, not a rebrand.
- Keeps your existing bans when the reference breaks them, and never imports the reference's numbers or claims.
- Plugs into Founder Ghostwriter as a mandatory style file for the account.

## Before you start

- Access to the reference posts (public permalinks or a signed-in browser)
- 3–10 of your own account's posts

## How to use Reference Style Model

1. **type.com:** Choose Use Skill to add Reference Style Model to your type.com workspace.
2. **Claude Code:** Download the skill and extract it into ~/.claude/skills/reference-style-model/ (or .claude/skills/reference-style-model/ in a project). Then ask Claude to use Reference Style Model.
3. **Codex:** Download the skill and extract it into ~/.agents/skills/reference-style-model/ (or .agents/skills/reference-style-model/ in a repository). Then ask Codex to use Reference Style Model.

## Frequently asked questions

### What is Reference Style Model?

Reference Style Model is an AI skill for sales & marketing and creative & design teams that you can use in Claude Code, Codex, or type.com. It works with LinkedIn.

### What does Reference Style Model do?

"We want our LinkedIn to sound like them." Turns a reference company's public posts into a style spec: numbered moves, each proved by a real line, and a pre-publish checklist. It transfers form, never their facts.

### What do I need to use Reference Style Model?

Access to the reference posts (public permalinks or a signed-in browser); 3–10 of your own account's posts.

### Can I use Reference Style Model in Claude Code or Codex?

Yes. Reference Style Model is a standard Agent Skills folder with a SKILL.md file. Download it and extract it into ~/.claude/skills/ for Claude Code or ~/.agents/skills/ for Codex.

### Who created Reference Style Model?

Barbara Jovanovic from Startup Cookie created Reference Style Model. The type.com team reviews every skill before it is published in the Skills Library.

## Example output

Example output from a sample run. Unless the example says otherwise, company names and figures are sample data.

````markdown
```
Style model: Example Co · LinkedIn company page
Reference: 19 public posts from a developer-tools company (Jan 2025–Oct 2026)
Gaps: feed is login-walled; posts collected one permalink at a time. Line breaks on 7 posts marked "inferred".

MOVES (excerpt)
3. Mechanism sentence with literal outputs
   Reference: "The policy engine returns allow, deny, or ask."
   For Example Co: "The scheduler returns a confirmed slot, a waitlist position, or a callback request."
4. Honest shipping status
   Reference: "It's in private beta. It handles Python today; Go is next."
   For Example Co: "Live for clinics in the US. EU data residency ships in Q1."
7. Labelled resource stack (launches only)
   Blog: <link>
   Docs: <link>

GAP ANALYSIS
Already doing: problem-led openers, "we" voice
New: mechanism sentences, shipping status, named credit, labelled link stack
Not applicable: specs block with benchmark numbers (no published benchmarks; Gate A would kill invented ones)

DELIBERATELY NOT COPIED
- "isn't X, it's Y" (banned in house style) → replacements: "X. Y as well." / "X, and Y."
- Em dashes, marketing adverbs ("seamlessly")
- Alarm framing ("🚨 Calling all...")

PRE-PUBLISH PASS (excerpt)
☐ 1. Opener names a reader problem in one clause
☐ 2. At least one mechanism sentence with literal outputs
☐ 3. Every number has a primary Example Co source
☐ 11. Still true and useful if the reference company didn't exist

Scope: Example Co company page only. Founder's personal posts unaffected.
Wired in: founder-ghostwriter Step 0 (mandatory load) and Gate D.
```
````

## Related skills

- [AI List Inclusion Outreach](https://type.com/library/skills/ai-list-inclusion-outreach.md): Get named on the "best X" lists that ChatGPT, Claude, Perplexity and Google AI answers already cite for your buyers' questions. Ranks the lists, drafts a three-line request for each, and tracks follow-ups.
- [Transparent Lower Third](https://type.com/library/skills/transparent-lower-third.md): On-brand animated name tags for testimonial videos, delivered as transparent ProRes 4444 files that drop straight onto a CapCut or Premiere timeline, with brand pulled from your site and a frame-by-frame check.
- [Event Post Fact-Check](https://type.com/library/skills/event-post-fact-check.md): Draft post-conference posts with every speaker, affiliation, session, sponsorship, photo and hashtag checked against the official program, keeping what the program proves separate from what someone remembers.
- [Social Batch Review Calendar](https://type.com/library/skills/social-batch-review-calendar.md): Replace PDFs and chat recaps with a Type app your client can actually review: a weekly calendar with each post's first paragraph and full image, copyable captions, downloadable visuals, and the approver on every card.
- [Founder Ghostwriter](https://type.com/library/skills/founder-ghostwriter.md): Ghostwrite LinkedIn and X posts in a founder's real voice, grounded only in things they actually said. Every post has a verbatim anchor, passes six gates, and is read by fresh-context reviewers before you see it.
- [Slop Edit](https://type.com/library/skills/slop-edit.md): Line-edit any draft so it stops reading as AI-written. A mechanical lint catches the surface patterns, a sentence-by-sentence pass fixes agentless and vague constructions, and every reviewer correction becomes a rule.

## Package files

- SKILL.md

## SKILL.md

````markdown
---
name: reference-style-model
description: Learn a company page's writing style, formatting and tone from a reference account's public posts, then turn it into a reusable style model with numbered moves and a pre-publish pass. Use when someone says "we want our LinkedIn to sound like <company>", "learn from these posts", or wants a consistent brand-page voice. Transfers form, never the reference's facts or numbers.
---

# Reference Style Model

Turns "make us sound like them" into a checkable spec. The output is a small skill (or a file loaded by your ghostwriting skill) that every draft for one account runs through.

**The governing rule: the moves transfer, the evidence does not.** The reference company's numbers, benchmarks, customers and claims never appear in your drafts. Copying their metric format without a real number of your own is a fabrication.

## Inputs
- The reference account URL (for example a LinkedIn company page).
- The target account, plus 3–10 of its existing posts.
- The target's existing bans (house style rules, slop-edit call-outs).

## Step 1. Harvest the reference corpus
- Company feeds are usually login-walled (403 or a sign-in page). Individual post permalinks are often publicly readable: collect them one at a time from search results, the company's site, or a signed-in browser the user controls.
- Aim for 15–25 posts across at least a year. Paste each **verbatim** with its URL and date.
- If your reader collapses newlines, say so and mark line breaks as **inferred**. Ask the user for 5–6 screenshots from their logged-in view to confirm formatting.
- **Scope to the posts.** Exclude the reference's marketing site and ads unless the user asks; site copy is a different register.

Save as `01-reference-corpus.md`, grouped into **house forms** (launch, research write-up, event, hiring, reshare-with-comment, argument/opinion, release notes, etc.).

## Step 2. Measure and compare
Build `02-form-inventory.md`:
- Metrics per form for the reference **and** the target's own posts: length, sentence length, emoji count and position, hashtags, links (body vs comment, labelled vs bare), questions, use of "we", named people, numbers per post.
- A **post-type → model** table: for each kind of post the target makes, which reference form to follow.
- **House tics**: small repeated habits (a structural emoji that opens launch posts, a `The specs:` block, a labelled link stack like `Blog: / Paper: / Docs:`).

## Step 3. Write the moves
In `SKILL.md` for the style model, write 8–12 **moves**. Each move:
- has a name,
- is proved by one verbatim line from the reference,
- says when to use it,
- shows a parallel line written for the target **with a target fact** (or `[needs real number]`).

Moves that commonly transfer well from strong technical brand pages:
1. Direct, problem-led opening (the problem the reader has, in their words).
2. Mechanism sentence with literal outputs ("returns allow, deny, or ask").
3. A specs block with real metrics, including cost, when real metrics exist.
4. Honest shipping status (alpha, waitlist, "works for X, not yet Y").
5. Named human credit (who built it, tagged).
6. Structural emoji used as a section marker, not decoration.
7. Labelled resource stack at the end of launches (`Blog:` / `Docs:` / `Paper:`).
8. Mechanism-led close on argument posts (end on how it works, not a CTA).
9. Reshare-with-comment: one line of real opinion on someone else's work.
10. A self-implicating or against-interest line where evidence supports it.

## Step 4. Keep the target's bans stronger than the reference
If the reference uses something the target has banned (em dashes, "isn't X, it's Y", marketing adjectives, alarm/FUD framing), **keep the ban** and give 2–3 parallel-positive replacements that get the same rhythm. Write these in a "Deliberately not copied" section with the reason.

## Step 5. Gap analysis
Compare the moves to the target's existing posts. Mark each move: **already doing it** / **new** / **not applicable**. Usually this is a small shift, not a rebrand: say so.

## Step 6. Write the pre-publish pass
A numbered checklist (aim for 10–15 points) that a draft must pass. Each point is a yes/no question. Include at minimum:
1. Does the opener name a problem the reader has, in one clause?
2. Is there at least one mechanism sentence with literal outputs?
3. Is every number from a primary source of the **target** company?
4. Is shipping status honest (no implied GA)?
5. Is anyone who did the work named and tagged?
6. Do emoji (if any) mark structure only?
7. Is there a labelled link stack (launch/release) or a mechanism close (argument)?
8. Zero banned constructions from the target's list?
9. Zero marketing adjectives/adverbs?
10. No named competitors, no headcount claims (unless the target allows them)?
11. Would this post still be true and useful if the reference company didn't exist?

## Step 7. Record default rulings and scope
- **Scope:** which account(s) this model applies to. A brand-page model never overwrites a person's personal voice, and doesn't apply to other platforms unless the user says so.
- **Default rulings** you made (for example "the 'no CTA' rule now allows the labelled link stack on launches") with a note that the user can override them.

## Step 8. Wire it in
Save as its own skill (or a file in the ghostwriting reference set). If a ghostwriting skill is in use, add this model as a **mandatory load at Step 0** for the scoped account and run its numbered pass as the **voice gate**, so it applies to every draft without being asked.

## Deliverable
```
reference-style-model/<target>-<platform>/
  SKILL.md                 # scope, moves (each with verbatim proof), not-copied list, pre-publish pass
  01-reference-corpus.md   # verbatim posts, URLs, dates, grouped by form; inferred line breaks marked
  02-form-inventory.md     # metrics vs target, post-type→model table, house tics
```
Report to the user: how the corpus was sourced and its gaps, what's new vs what they already do, what was deliberately not copied, default rulings, and where it's wired in.
````
