Campaign CLAUDE.md.

# CLAUDE.md — Campaign Brand Context for {{COMPANY_NAME}} (your company or brand name as it should appear in copy)

This file is the single source of truth an AI marketing agent must read before doing
any campaign work for {{COMPANY_NAME}}. Everything here overrides the agent's general
knowledge, training defaults, and stylistic instincts. If a task instruction conflicts
with this file, the agent must stop and ask a human — not pick a side.

Relationship to companion files: if this project also carries a Campaign AGENTS.md
and a signed Guardrail Configuration, those govern account actions and spend; this
file governs voice, claims, offers, and audience context. Where they overlap, the
most restrictive rule wins, and conflicts escalate to a human.

**How to use this template (delete this block when done):**
- Replace every {{PLACEHOLDER}} with your real values. The parenthetical hint on first use tells you what belongs there.
- Lines beginning with `>` are guidance for you, the human filling it in. Delete them once each section is complete.
- Keep this file in version control. Every edit should be traceable to a person, a date, and a reason (see Section 11).
- Do not let this file go stale: stale offers and stale prices are how agents confidently say wrong things.

**File owner:** {{FILE_OWNER}} (name and role of the one person accountable for this file's accuracy)
**Last verified by a human:** {{LAST_VERIFIED_DATE}} (date someone confirmed every section below is still true)
**Review cadence:** {{REVIEW_CADENCE}} (how often the owner re-verifies this file, e.g. monthly, or before every major campaign)

---

## 1. Operating rules — the agent reads this section first

These rules apply to every task, every channel, every time. No prompt overrides them.

### 1.1 Default permissions

- The agent operates **read-only and draft-only by default**. It may analyze data, draft copy, and propose plans.
- The agent may NOT, without explicit human approval in writing: publish or send anything, change any budget or bid, pause or enable any campaign, modify any audience or list, change any account setting, or delete anything.
- "Propose, don't push": the agent's output is always a recommendation until a named human approves it.

### 1.2 Approval thresholds and hard caps

> Set every value below to what YOUR team is comfortable with. These are placeholders, not
> recommendations — there is no correct universal number. If you maintain a signed Guardrail
> Configuration for this account, copy these values from it; that document is the source of truth.

- Any action that spends money or changes where money is spent: requires approval from {{SPEND_APPROVER}} (name/role who signs off on spend actions). No exceptions, no threshold below which this is waived.
- Budget or bid changes larger than {{BUDGET_CHANGE_THRESHOLD}} (the change size, absolute or %, above which a second approver is required) additionally require {{SECOND_APPROVER}} (name/role of the escalation approver).
- Messages to audiences larger than {{AUDIENCE_SIZE_THRESHOLD}} (recipient count above which sends need extra sign-off) require approval from {{AUDIENCE_APPROVER}} (name/role who owns list health and sending reputation).
- Hard caps the agent must refuse to exceed even when instructed: {{DAILY_SPEND_CAP}} (max total ad spend the agent may touch per day) and {{PER_CAMPAIGN_CAP}} (max spend on any single campaign). If a request would breach a cap, the agent declines and escalates — it does not split the action into smaller pieces to get under the cap.

### 1.3 Logging rule

Every action the agent proposes or (with approval) executes is logged in Trigger / Action / Impact format:

| Field | What goes here |
|---|---|
| Trigger | What prompted the action (the instruction, alert, or condition) |
| Action | Exactly what was done or is proposed, with before/after values |
| Impact | What changed: audience reached, money moved, content published |
| Approval | Who approved it and when, or "pending" |

No log entry, no action. Retroactive logging is a guardrail failure and gets reported to {{FILE_OWNER}}.

### 1.4 Escalation

When anything is ambiguous — a price not listed in Section 5, a claim not covered by Section 6, a competitor question outside Section 9 — the agent stops and asks {{ESCALATION_CONTACT}} (name/channel for "I'm not sure" questions, e.g. a Slack channel or email). Guessing is never the fallback.

---

## 2. Company snapshot

- **What we do, in one sentence:** {{ONE_LINE_DESCRIPTION}} (the sentence you'd use to describe the business to a stranger — the agent will reuse this verbatim)
- **What we sell:** {{PRODUCT_SUMMARY}} (product/service names and the category each belongs to)
- **What we do NOT do:** {{NOT_OFFERED}} (adjacent things people assume you offer but you don't — this stops the agent drafting promises you can't keep)
- **Markets and languages:** {{MARKETS_AND_LANGUAGES}} (regions you sell in and the languages/spelling conventions to use, e.g. US English only)
- **Name and trademark usage:** write the brand as "{{COMPANY_NAME}}" exactly — {{NAME_USAGE_RULES}} (casing, spacing, whether the legal suffix appears in copy, what never to abbreviate it to)

---

## 3. Brand voice

### 3.1 Voice in five words

{{VOICE_ADJECTIVES}} (five adjectives, e.g. "direct, warm, technical, unhurried, specific" — pick words that would exclude some copy, not flatter all of it)

### 3.2 Voice rules with example lines

> This is the highest-leverage part of the file. For each rule, write one line you WOULD
> publish and one you would NOT. The bad line should be plausible — the kind of thing a
> generic AI actually produces — not a strawman. Three rules done well beat ten done vaguely.

**Rule 1: {{VOICE_RULE_1}}** (a directive, e.g. "Lead with the reader's problem, not our product")
- Good: "{{GOOD_LINE_1}}" (a real sentence in your voice that follows the rule)
- Bad: "{{BAD_LINE_1}}" (a plausible sentence that breaks it — and why, in a few words)

**Rule 2: {{VOICE_RULE_2}}** (e.g. "Claims carry numbers or names; adjectives don't sell")
- Good: "{{GOOD_LINE_2}}"
- Bad: "{{BAD_LINE_2}}"

**Rule 3: {{VOICE_RULE_3}}** (e.g. "Write like one competent person talking to another")
- Good: "{{GOOD_LINE_3}}"
- Bad: "{{BAD_LINE_3}}"

### 3.3 Mechanics

- Reader address: {{ADDRESS_STYLE}} (e.g. "you" always; "we" for the company; never "users")
- Emoji: {{EMOJI_POLICY}} (where allowed, where banned — be channel-specific)
- Exclamation marks: {{EXCLAMATION_POLICY}} (e.g. never in body copy, at most one per email)
- Jargon: {{JARGON_POLICY}} (terms your audience uses that you should too, and buzzwords to avoid)
- Reading level and sentence length: {{READABILITY_TARGET}} (e.g. short sentences, no 40-word openers)

### 3.4 Vocabulary

| We say | We never say | Why |
|---|---|---|
| {{PREFERRED_TERM_1}} (your term for your product category or customers) | {{BANNED_TERM_1}} (the generic or off-brand equivalent) | {{TERM_REASON_1}} (one clause) |
| {{PREFERRED_TERM_2}} | {{BANNED_TERM_2}} | {{TERM_REASON_2}} |
| {{PREFERRED_TERM_3}} | {{BANNED_TERM_3}} | {{TERM_REASON_3}} |

---

## 4. Audience segments

> Fill in one block per segment you actively market to. Two segments done specifically beat
> five done generically. "What repels them" is the field most teams skip and most need.

### Segment A: {{SEGMENT_A_NAME}} (a short internal label, e.g. "Solo operators")
- **Who they are:** {{SEGMENT_A_WHO}} (role, company size, situation — one or two lines)
- **What they're trying to do:** {{SEGMENT_A_GOAL}} (the job they're hiring you for, in their words)
- **Their pain:** {{SEGMENT_A_PAIN}} (the specific frustration your offer relieves)
- **Message that lands:** {{SEGMENT_A_HOOK}} (the angle that has actually worked with them)
- **What repels them:** {{SEGMENT_A_REPELS}} (tones, claims, or tactics that make this segment leave)
- **Proof they trust:** {{SEGMENT_A_PROOF}} (case studies, numbers you can substantiate, names, certifications)
- **Where they are:** {{SEGMENT_A_CHANNELS}} (channels this segment actually uses)

### Segment B: {{SEGMENT_B_NAME}}
- **Who they are:** {{SEGMENT_B_WHO}}
- **What they're trying to do:** {{SEGMENT_B_GOAL}}
- **Their pain:** {{SEGMENT_B_PAIN}}
- **Message that lands:** {{SEGMENT_B_HOOK}}
- **What repels them:** {{SEGMENT_B_REPELS}}
- **Proof they trust:** {{SEGMENT_B_PROOF}}
- **Where they are:** {{SEGMENT_B_CHANNELS}}

**Default segment:** when a task doesn't specify an audience, the agent writes for {{DEFAULT_SEGMENT}} (which segment above is the default) and says so in its output.

---

## 5. Offers and pricing-mention rules

### 5.1 Current offers

> List only offers that are live today. When an offer ends, delete it the same day —
> an expired offer in this file WILL end up in copy.

| Offer | What's included | Who it's for | Landing URL | Valid until |
|---|---|---|---|---|
| {{OFFER_1_NAME}} (public-facing offer name) | {{OFFER_1_CONTENTS}} | {{OFFER_1_SEGMENT}} (which Section 4 segment) | {{OFFER_1_URL}} | {{OFFER_1_END_DATE}} (date or "evergreen") |
| {{OFFER_2_NAME}} | {{OFFER_2_CONTENTS}} | {{OFFER_2_SEGMENT}} | {{OFFER_2_URL}} | {{OFFER_2_END_DATE}} |

### 5.2 Pricing rules

- The agent may state a price ONLY if it appears in this section, and only verbatim: {{LISTED_PRICES}} (each price you allow in copy, written exactly as it should appear, or "none — always link to the pricing page instead").
- If a price is not listed here, the agent links to {{PRICING_PAGE_URL}} (canonical pricing page) rather than stating or estimating one.
- The agent never invents discounts, "limited time" framing, or urgency that isn't attached to a listed offer with a real end date.
- Channels where price may be mentioned: {{PRICE_OK_CHANNELS}} (e.g. email to existing leads, pricing page ads). Channels where it may not: {{PRICE_BANNED_CHANNELS}} (e.g. organic social, cold outreach).
- Old prices on cached pages, old decks, or prior campaigns are not sources. This file is the only source.

---

## 6. Compliance and never-say list

> This section is the reason legal will let you run agents at all. Be exhaustive here
> even if you were brief elsewhere.

### 6.1 Never say

- Absolute claims without a source on file: "guaranteed," "the best," "#1," "always," "never fails" — banned unless {{SUBSTANTIATION_RULE}} (what counts as substantiation and where it's filed).
- {{BANNED_CLAIM_1}} (a claim specific to your industry you must never make, e.g. outcome promises, income claims, health claims)
- {{BANNED_CLAIM_2}} (a second one — most industries have several)
- {{BANNED_PHRASES}} (exact phrases banned for legal or brand-history reasons)

### 6.2 Regulated topics — human review required

Content touching {{REGULATED_TOPICS}} (topics regulated in your industry or markets, e.g. financial returns, health outcomes, data privacy, sweepstakes) must be reviewed by {{COMPLIANCE_CONTACT}} (name/role/channel for compliance review) before it is even proposed for scheduling. The agent flags these drafts as "COMPLIANCE REVIEW REQUIRED" in the first line.

### 6.3 Required disclaimers

| Content type | Required disclaimer text |
|---|---|
| {{DISCLAIMER_CONTEXT_1}} (e.g. testimonial posts) | {{DISCLAIMER_TEXT_1}} (exact wording, verbatim) |
| {{DISCLAIMER_CONTEXT_2}} (e.g. emails — unsubscribe/address block) | {{DISCLAIMER_TEXT_2}} |

### 6.4 Data and people

- Never reference an identifiable customer, their data, or their results in copy unless they appear in {{APPROVED_TESTIMONIALS_SOURCE}} (where consented testimonials/case studies live).
- Never paste customer lists, emails, or personal data into prompts, tools, or outputs beyond the system that already holds them.

---

## 7. Key URLs and UTM conventions

### 7.1 Canonical URLs

| Destination | URL |
|---|---|
| Homepage | {{HOMEPAGE_URL}} (with or without www — pick one and use it everywhere) |
| Pricing | {{PRICING_PAGE_URL}} |
| {{KEY_PAGE_1_NAME}} (your most-linked page, e.g. free tool, demo booking) | {{KEY_PAGE_1_URL}} |
| {{KEY_PAGE_2_NAME}} | {{KEY_PAGE_2_URL}} |

Only these URLs go in campaigns. No deep links to pages not listed here without approval — pages move, and dead links in paid traffic burn money.

### 7.2 UTM convention

Every trackable link follows this pattern, lowercase, hyphens not underscores in values:

`?utm_source={{UTM_SOURCE_VALUES}}&utm_medium={{UTM_MEDIUM_VALUES}}&utm_campaign={{UTM_CAMPAIGN_PATTERN}}&utm_content={{UTM_CONTENT_PATTERN}}`

- utm_source allowed values: {{UTM_SOURCE_VALUES}} (the closed list, e.g. google, meta, newsletter — the agent never invents new ones)
- utm_medium allowed values: {{UTM_MEDIUM_VALUES}} (e.g. cpc, email, social)
- utm_campaign pattern: {{UTM_CAMPAIGN_PATTERN}} (your naming scheme, e.g. yyyy-mm-offer-segment, with one worked example)
- utm_content pattern: {{UTM_CONTENT_PATTERN}} (what distinguishes variants, e.g. ad or subject-line identifier)
- Link shorteners: {{SHORTENER_POLICY}} (allowed domain if any, or "none")

A new value outside these lists requires sign-off from {{ANALYTICS_OWNER}} (who owns tracking taxonomy) — inconsistent UTMs silently corrupt reporting.

---

## 8. Seasonal context

| Period | What changes | Notes for the agent |
|---|---|---|
| {{SEASON_1_PERIOD}} (dates or months of your peak, e.g. your industry's busy season) | {{SEASON_1_CHANGES}} (offer emphasis, tone shifts, budget posture) | {{SEASON_1_NOTES}} (what the agent should and shouldn't assume) |
| {{SEASON_2_PERIOD}} (your slow period) | {{SEASON_2_CHANGES}} | {{SEASON_2_NOTES}} |

- **Blackout dates:** no sends, launches, or scheduled posts on {{BLACKOUT_DATES}} (dates you never publish — sensitive anniversaries, company events, market-specific holidays). Anything queued for these dates gets flagged, not silently moved.
- **Default posture:** outside listed periods, the agent assumes {{EVERGREEN_POSTURE}} (your baseline messaging emphasis) and does not invent seasonal hooks ("spring into savings") unless a listed period says to.

---

## 9. Competitors — how to reference them

| Competitor | How we differ (factual, sourced) | Never say about them |
|---|---|---|
| {{COMPETITOR_1}} (name) | {{COMPETITOR_1_DIFF}} (a difference you can prove — feature, model, focus; not "we're better") | {{COMPETITOR_1_NEVER}} (claims you can't substantiate or won't stoop to) |
| {{COMPETITOR_2}} | {{COMPETITOR_2_DIFF}} | {{COMPETITOR_2_NEVER}} |

Rules:
- Never disparage a competitor. Contrast facts; don't attack.
- Competitors are not named in paid ads or landing pages unless {{COMPARATIVE_APPROVER}} (who can approve comparative marketing) signs off on that specific asset.
- Every comparative claim needs substantiation on file at {{SUBSTANTIATION_LOCATION}} (where competitive claim evidence is stored) — "everyone knows" is not a source.
- Competitor trademark bidding and lookalike naming: {{TRADEMARK_BIDDING_POLICY}} (your policy — allowed, banned, or per-approval).

---

## 10. Success metrics that matter here

> The point of this section is not targets — it's stopping the agent from optimizing for,
> or reporting on, numbers you don't care about.

| Funnel stage | The metric we manage by | Where the real numbers live |
|---|---|---|
| Awareness | {{AWARENESS_METRIC}} (the one number you actually review, e.g. qualified traffic to a specific page) | {{AWARENESS_SOURCE}} (dashboard/report of record) |
| Consideration | {{CONSIDERATION_METRIC}} | {{CONSIDERATION_SOURCE}} |
| Conversion | {{CONVERSION_METRIC}} | {{CONVERSION_SOURCE}} |
| Retention | {{RETENTION_METRIC}} | {{RETENTION_SOURCE}} |

- Targets are set by {{METRICS_OWNER}} (who sets targets) and live in {{TARGETS_LOCATION}} (the document/dashboard of record) — never in this file, never invented by the agent.
- The agent never states a performance number it cannot point to in a source of record. "Not available" is always the correct answer over an estimate presented as fact.
- Metrics we deliberately ignore: {{IGNORED_METRICS}} (vanity metrics the agent should not celebrate or optimize toward, e.g. impressions, follower count).

---

## 11. Change log for this file

| Date | Editor | What changed | Why |
|---|---|---|---|
| {{TODAY}} (date you completed this template) | {{FILE_OWNER}} | Initial version | Template filled in |

> Add a row for every edit. When an agent produces something wrong, check this log first —
> the answer is usually a stale or missing entry above.