Launch Campaign

Launch a Fern paid acquisition campaign

You are scaffolding a new campaign folder under campaigns/<slug>/ that drives traffic to a single landing page. Output: a complete campaigns/<slug>/ tree (brief + copy + creative briefs + results scaffolding) plus a prefix entry in campaigns/README.md. Reference shape: campaigns/leave-stainless/ and campaigns/agent-ready/.

Step 0 — Establish the spine

If the user hasn't already given you all of this, ask in ONE short message:

1. Slug — kebab-case folder name (e.g. agent-ready, leave-stainless). This is also the utm_campaign value.
2. LP path — usually /lp/<slug>. Confirm the page exists at src/app/lp/<slug>/page.tsx.
3. One-line wedge — the angle that distinguishes this campaign from the others (e.g. "offense on the invisible-audience problem", "defense against a competitor wind-down", "side-by-side evaluation").
4. Anything stable about audience or proof points they want locked in (named customers, specific quotes, a stat).

That's it. Don't pepper them. Derive everything else from the LP page and the AGENTS.md writing style.

Step 1 — Read the source of truth

Before writing any campaign file, read in parallel:

• src/app/lp/<slug>/page.tsx — for the actual headline, named features, anchors (every id= in the JSX), and CTA destinations. The campaign copy must echo the LP's language, not invent new claims.
• AGENTS.md — the "Writing style" section. Every line of ad copy obeys it.
• campaigns/README.md — for the canonical variant-ID scheme and the list of registered prefixes (so the new prefix doesn't collide).
• campaigns/leave-stainless/ and campaigns/agent-ready/ — structural reference for the new files. Match section ordering and header style exactly so the campaigns stay diff-comparable.
• Any blog post the LP cross-links to (e.g. content/blog/agent-score.mdx) — for verbatim customer quotes, named stats, and proof points.

Step 2 — Pick the prefix

Uppercase initials of the slug, with a trailing dash. leave-stainless → LS-. agent-ready → AR-. fern-vs-speakeasy → FVS-. If the natural prefix collides with one already registered in campaigns/README.md, pick the next-shortest distinguishing letters and confirm with the user.

Step 3 — Build the tree

Create exactly this structure (use the _template/ directory as the placeholder shape, but fill every field from the LP and the user's wedge — do not copy and paste placeholders forward):

campaigns/<slug>/
├── brief.md
├── copy/
│   ├── linkedin.md
│   └── google.md
├── creative/
│   ├── README.md
│   ├── <PFX>-LI-SI-01.md          # Lead hook
│   ├── <PFX>-LI-SI-02.md          # Tool / product hook
│   ├── <PFX>-LI-SI-03.md          # Social-proof / testimonial hook
│   ├── <PFX>-LI-SI-04.md          # Data / stat hook
│   ├── <PFX>-LI-CONV-01.md        # Conversation ad banner
│   ├── <PFX>-GD-LS-01.md          # Google Display landscape
│   ├── <PFX>-GD-SQ-01.md          # Google Display square
│   └── exports/.gitkeep
└── results/
    ├── README.md
    ├── performance.csv            # Header row only
    └── learnings.md

Four LinkedIn single-image variants is the proven shape. Don't ship fewer (the campaign needs hook diversity for two-week iterations) and don't ship more (more variants split spend before you have signal).

Step 4 — Brief (brief.md)

Sections in order, matching campaigns/leave-stainless/brief.md:

1. Landing page — URL, full UTM scheme, and a list of every anchor on the LP worth deep-linking (one per id= from the JSX).
2. Audience — primary, secondary, tertiary. Then LinkedIn targeting (job titles, company filters, group signals, account audiences) and Google targeting (branded terms, category terms, adjacent intent, negative keywords). All keywords lowercase.
3. Goals — primary KPI (almost always booked demos via BookDemoForm), secondary KPIs, budget TBD, window TBD.
4. Value props (priority order) — 3 to 5, drawn from the LP and the user's wedge.
5. Proof points to lean on — customers, stats, quotes, certifications. Pull verbatim from the LP and the cross-linked blog post.
6. Anti-patterns to avoid — at minimum: em-dashes anywhere; "AI-powered" without naming the feature; "10x faster" without a source; future hedging; disparaging competitors by name. Add campaign-specific anti-patterns when relevant.

Step 5 — LinkedIn copy (copy/linkedin.md)

Four single-image hooks, two text ads, one conversation ad. Limits at the top of the file (Single Image intro ≤600 max but aim ≤200; headline ≤200 but aim ≤70; description ≤100; text-ad headline ≤25; text-ad description ≤75; conversation ad opening ≤400). For every variant include intro, headline, description, full destination URL with UTMs, and a markdown link to the paired creative brief.

The four single-image hooks should each play a different beat:

• SI-01 — Lead hook — the campaign's strongest one-sentence wedge. Urgency or framing.
• SI-02 — Tool / product hook — the feature or tool that makes the campaign land. Pair with the deepest LP anchor.
• SI-03 — Social-proof hook — a verbatim customer quote with monochrome customer logo. No background imagery; solid brand-green field.
• SI-04 — Data hook — one big statistic, one secondary line, one CTA. The number does the work.

Print character counts inline (e.g. Intro (181 chars):) so they're easy to QA. Verify each one before committing.

Step 6 — Google copy (copy/google.md)

Two RSA ad groups (a general one and one keyword-tilted to the strongest deep-link), 15 headlines (≤30) and 4 descriptions (≤90) per RSA, sitelinks, callout extensions (≤25 each), structured snippets, and Google Display short headlines / long headline / descriptions. Format as tables with character counts so the QA pass is fast.

Step 7 — Creative briefs (creative/<PFX>-*.md)

Each brief is three sections, matching campaigns/leave-stainless/creative/:

1. Spec — platform, slot, dimensions, file path, paired copy variant link, paired UTM.
2. Technical asset outline — either a Mermaid diagram or a line-diagram showing the actual layout (boxes for headline, logo strip, CTA chip, customer-logo row). Concrete enough that a designer can rebuild without ambiguity. Then a bullet list calling out colors, type sizes, separators.
3. Background imagery — Midjourney — one prompt in the house Hiroshi Yoshida template. Pick a scene that maps to the variant's emotional read. Close the brief with a one-line "Emotional read:" callout.

The house Midjourney template:

[scene description in Hiroshi Yoshida landscape style], atmospheric [weather], long-exposure smoothing motion, muted earthy colors, tactile analog grain --v 7.0 --p xbmuohp --raw --exp 100 --ar {ratio}

Brand asset rules locked across all briefs:

• Fern wordmark at /public/fern-wordmark.svg, leaf mark at /public/fern-leaf.svg.
• Customer logos at /public/customers/<slug>/<slug>-monochrome.svg against imagery or colored fields.
• Brand green #008700 (light) / #9CDD8D (dark).
• Wordmark top-left at 64px equivalent on a 1200-wide canvas, 5% margin.
• Headline ~56pt on a 1200-wide canvas.
• For score readouts or product mocks, render as clean type, not a screenshot of the live product. The live UI changes; the ad shouldn't.

Aspect ratios by slot are documented in the campaign's creative/README.md. Copy that table verbatim from campaigns/leave-stainless/creative/README.md.

Step 8 — Results (results/)

Three files:

• README.md — explain how to log, list data sources (LinkedIn Campaign Manager, Google Ads, GA4), document the CSV schema, and describe the two-week review cadence (sort by cost-per-booking, sort by CTR, update learnings).
• performance.csv — header row only: date,platform,campaign,variant_id,impressions,clicks,ctr,spend,landing_page_views,demo_bookings,cost_per_booking,notes. Add a campaign-specific secondary conversion column (e.g. agent_score_runs) if the LP has one.
• learnings.md — empty "Working" and "Cutting" sections, then a "Test next" list of pre-launch hypotheses worth queuing (audience split, CTA copy split, destination test, background-imagery test, etc.).

Step 9 — Register the prefix

Append two lines to campaigns/README.md:

1. A row in the "Campaign | Prefix | Example" table.
2. A bullet under "Active campaigns" describing the LP and the angle.

Step 10 — Verify

Before reporting done, run these checks in parallel:

• find campaigns/<slug> -type f | sort — matches the file tree above exactly.
• grep -r "utm_content=<PFX>-" campaigns/<slug> — every variant ID is unique and registered.
• grep -rE "(^|[^-])—([^-]|$)" campaigns/<slug>/copy — confirms no em-dashes in the ad copy. (Em-dashes in section headers like ### <PFX>-LI-SI-01 — Headline are the project convention and OK; em-dashes inside intro / headline / description fields are not.)
• grep -rE "AI-powered|is designed to" campaigns/<slug> — no future hedging or unnamed AI features in the copy. (Anti-pattern lines in brief.md are expected.)
• Eyeball every LinkedIn intro ≤600 chars, headline ≤200, description ≤100, text-ad headline ≤25, text-ad description ≤75, conversation opener ≤400.
• Eyeball every Google RSA headline ≤30, description ≤90.
• The variant-ID table in campaigns/README.md includes the new prefix row.

Step 11 — Stop

Do not commit. Do not push. Do not open a PR. Show the user the file list and ask whether to commit. They will iterate on copy before launch and prefer to bundle that with the first commit.

Anti-patterns when running this skill

• Don't invent customer quotes. Pull them verbatim from blog posts, case studies, or the LP. If a campaign needs a quote and none exists in the repo, surface that gap to the user instead of fabricating one.
• Don't invent stats. If the LP says "73% of top API docs sites ship without an llms.txt", that's the only stat you have. Don't extrapolate.
• Don't restructure the existing campaign layout to "improve" it. Other campaigns depend on the shape being diff-comparable.
• Don't ship a campaign with fewer than four LinkedIn single-image hooks. The two-week iteration loop needs hook diversity.
• Don't deep-link to an anchor that doesn't exist on the LP. Grep the page for id= before linking.