Google Ads Campaigns as Code

Google Ads search campaigns as reviewable JSON specs — offline plan, --apply creates PAUSED; opt-in budget-cap script pauses scoped campaigns. Use when an agent builds a campaign.

Google Ads Search Campaigns as Code

The only automatic action tolerable on a system that spends money is the one that PREVENTS spending. Everything else is a proposal a human reads before it runs.

A campaign is a versioned artifact: an agent writes the spec, a human reads the diff, one command creates it paused. Reference implementation: Google Ads API (google-ads-python). The shape is vendor-agnostic; "vendor" below means whichever ad network you point it at.

Access, data, and network

Item	Reality
Default mode	Plan only (no flag, or --dry-run) — no client built, no environment variable read at all, no network call
Mutating mode	Only with the explicit --apply flag, after a human go on the plan. Omitting a flag can never create anything
Dry-run output	Carries no account id and no credential. Safe to paste into a ticket for review
Credentials	Env vars, read in load_env() on the mutating path only. Never hardcoded, never printed, never logged
Persisted by ads-search.py	Nothing. Specs are files you already own; the script writes no state
budget-cap-guard.js — live-account, opt-in, installed by a human	Runs hourly in the ad account's own script runner, not in the agent. Reads month spend; at the cap it pauses only ENABLED campaigns whose name contains the required NAME_CONTAINS prefix, and only once a human sets PAUSE_ON_CAP: true (shipped false). There is no account-wide mode. The agent never installs, edits, or schedules it
Written by budget-cap-guard.js	Creates one tab it owns (budget-cap-guard) in a dedicated spreadsheet you own, appends one row per hourly breach: timestamp, status, spend, cap, action, writer marker. No PII, no credential. Refuses to write into a tab that already exists with any other header
Retention (enforced, not announced)	Every run deletes rows older than CONFIG.RETAIN_DAYS (default 400) — only rows carrying its own writer marker, only in its own tab. 0 = keep forever, an explicit opt-out
Invalid config	Missing cap, prefix, or sheet URL, a wrong type, or unreadable spend: the run throws before any pause or write and the script runner reports the failure. Never falls back to "every campaign"
Leaves the machine	Only on the mutating path: the campaign tree, to the ad network you authenticated against
Spend authority	None. Every campaign is created PAUSED. Going live is a human click in the vendor UI
Prerequisites (mutating path)	A dedicated virtualenv with pip install "google-ads==33.0.0" (the script refuses --apply on any other version). No developer token — sunset 2026-09-09. API access = the level of the Google Cloud project that issued the OAuth client: Test (test accounts only) · Explorer (production, 2,880 ops/day, no KeywordPlanIdeaService) · Basic (15,000 ops/day, requires the project's brand verification) · Standard (unlimited, ~10 business days review). Upgrade: Cloud console → Google Ads API overview → Upgrade access level. A new project is at Test and cannot write to production. OAuth Desktop client + refresh token generated elsewhere; an app left in Testing status issues refresh tokens that expire after 7 days. Manager (MCC) account only if you reach the account through one
Your obligations	Vendor ad policies, consumer-protection and advertising law apply to everything you publish. Regulated verticals (health, finance, legal, licensed trades) carry statutory constraints on top of vendor policy — clear the copy with the accountable human, not with this skill

Setup for the mutating path only: python3 -m venv .venv && .venv/bin/pip install "google-ads==33.0.0", then export GOOGLE_ADS_CLIENT_ID, GOOGLE_ADS_CLIENT_SECRET, GOOGLE_ADS_REFRESH_TOKEN, GOOGLE_ADS_CUSTOMER_ID, plus GOOGLE_ADS_LOGIN_CUSTOMER_ID only when access goes through a manager account. That is the whole list: the script declares no env var it does not consume. Conversion goals are an account-level setting — configure them in the vendor UI before you enable. The dry-run needs none of this.

When to Use

Trigger	Action
"build me a search campaign for X"	Write <segment>.json spec → dry-run → hand the plan to a human
"review the campaign before launch"	--dry-run, paste the output, wait for an explicit go
"create it / it's approved"	Run with --apply. Lands PAUSED. Say so out loud
"make it live" / "unpause"	Refuse. Point to the vendor UI. This skill has no enable path
"raise the budget"	Propose a spec diff. Never mutate budget on your own authority
"we're overspending"	Budget cap script (§10), installed and scoped by a human — not the agent

1. The spec is the artifact

One JSON file per segment. One segment = one campaign = one ad group cluster = one landing page.

Field	Type	Notes
campaign	string	Naming convention <segment>_<region>_<channel>_<offer>. Also the idempotence key. Validated against [A-Za-z0-9_.\-]{1,120} in code
segment	string	Slug, matches the file name
final_url	url	The segment's landing page. Ad copy and LP must say the same thing
currency	string	Descriptive only. Amounts are in the account's currency; nothing converts FX
geo.country_code	string	Required with any geo name. Zone names resolve within one country (§6)
geo.locale	string	Language of the zone names. Required with any geo name, no default
geo.include / geo.exclude	string[]	Plain zone names, resolved read-only at apply time (§6)
geo.worldwide	bool	The only way to target everywhere. Requires geo.note. Omitting geo.include does not mean worldwide — it aborts (§6)
geo.note	string	Why these zones. Reviewers read this first
language	string	Required, no default. english / french / spanish / german, or set language_constant (digits) for any other. Unknown values abort
bidding	string	MaximizeConversions; add target_cpa only once there is data
monthly_budget	number	The number the client agreed to
daily_budget	number	= monthly / 30.4. Not /30, not /31
budget_share	string	Share of the media budget, with a written rationale
ad_schedule.account_timezone	string	Windows are in the account timezone (§6)
ad_schedule.windows[]	object[]	days[], start, end as HH:MM
ad_groups[].keywords	object	exact[] / phrase[] / broad[], syntax [exact], "phrase", broad
ad_groups[].rsa	object	headlines[] 3–15, descriptions[] 2–4, paths[] ≤2. Both bounds warn in the dry-run; over-max is truncated on create, so the plan says how many will be dropped (§6)
negative_keywords	string[]	Campaign-level negatives (§8)
extensions	object	Sitelinks, callouts, structured snippets, call. Reviewed here, created in a later pass

Complete worked spec: campaign.example.json (acme-widgets_us_search_demo, 300/month → 9.87/day, 44 negatives, deliberately unrounded schedule). Its US/English values are illustrative; nothing in the code defaults to them.

Every field is type-checked before planning: a null, a string where a list belongs (list("United States") is 13 one-letter zones), a boolean budget (float(True) is 1.00), a 25:00 window — the plan aborts and lists every problem. Refuse, never coerce.

Budget split across segments needs a rationale in the file, not in a chat message: "acme-core 45% — biggest volume and strongest intent; acme-satellite 20% — narrower market, fewer zones." A share without a reason is a number nobody can review.

2. Sizing

The dry-run checks syntax, not sense. These are the thresholds it cannot enforce for you.

Question	Answer	Why
Keywords per ad group	5–20, one intent	The RSA must be relevant to all of them. 50 keywords = one of them is lying
Ad groups per campaign	1–5	They share the budget. More groups = each starves
Floor for MaximizeConversions	~15–30 conversions/month	Below that, smart bidding has no signal and burns budget learning
When to add target_cpa	After ~30 conversions, never at launch	A tCPA with no data throttles delivery to nothing
Broad match at launch	No	See §8

3. The dry-run is offline by construction

bash

python3 ads-search.py --spec campaign.example.json --dry-run
# Output (real, from a shell with no credentials and no SDK; elided for length):
# [ads-search] DRY-RUN — plan only. No client built, no env read, no network.
#   account            : (GOOGLE_ADS_CUSTOMER_ID, read at apply time — not here)
#   campaign           : acme-widgets_us_search_demo
#   channel            : SEARCH (content network OFF; search partners OFF)
#   STATUS             : PAUSED  <-- forced in code; no flag can change it
#   daily budget       : 9.87 -> 9870000 micros   (monthly 300)   [USD, no FX conversion]
#   geo include (1)   : United States
#     resolved within  : country US, locale en
#       - SATURDAY  10:00 -> 24:00
#       note: 10:07-23:59 rounded -> 10:00/24:00 (quarter-hour grid)
#   … (ad groups, keywords by match type, RSA limit checks elided)
#   campaign negatives (44) : free, cheap, price, prices, pricing, cost, how much, ...
# [ads-search] END DRY-RUN — nothing written.

The plan (no flag, or --dry-run) returns before build_client() is ever called. It reads no environment variable at all — load_env() is the only place this script touches the environment, and it sits past the dry-run return. The SDK import is behind a tolerant try/except, so the plan renders with google-ads not installed.

This is architecture, not a convenience flag. A full campaign — every keyword, every headline, every rounded slot, every micro — is reviewable by someone with no secret provisioned at all, and the output holds no account id, so it is safe to paste where the review happens. Keep the offline path pure: the moment a "harmless" env read sneaks in, you have re-coupled review to secrets and started printing identifiers into chat logs.

4. One atomic mutate, or nothing

The whole tree is wired with negative temporary resource names and sent as a single mutate:

python

budget_rn   = f"customers/{cid}/campaignBudgets/-1"   # placeholders resolved server-side
campaign_rn = f"customers/{cid}/campaigns/-2"
c.campaign_budget = budget_rn        # -2 points at -1 inside the same request
g.campaign        = campaign_rn      # ad group -3 points at -2 ...
svc.mutate(customer_id=cid, mutate_operations=ops)   # all of it, or none of it

Order: budget → campaign → criteria (language, geo, schedule, negatives) → ad group → keywords → responsive search ad.

Why this is the most expensive gotcha in the skill. Create the tree in sequential calls and any failure halfway leaves a campaign with a budget and no ad — an orphan. Orphans are not inert: a campaign object with a budget attached is a live spending surface the moment anyone enables it, and it will not show up in "campaigns I meant to create." One mutate means one rollback boundary. Never "just retry the last step."

5. PAUSED is a constant, not an instruction

python

c.status = enums.CampaignStatusEnum.PAUSED   # no flag, no env var, no argument reaches this

There is no --enable. Enabling is a human action in the vendor UI, by the person whose card is charged. A guardrail that matters is on the left side of an assignment, not in a prompt.

Ad groups are created ENABLED on purpose: the paused campaign gates delivery, so the tree is ready to serve the instant a human flips exactly one switch.

6. API traps

Trap	Root cause	Fix
Ads appear on unrelated sites	Network flags default permissive	Set all four: search ON, target_content_network=False, target_partner_search_network=False, partners off unless reviewed
INVALID_ARGUMENT on geo	Zone names are not resource names	Resolve names → geo target constants with a separate READ-ONLY call before the mutate. Never resolve inside it
Campaign silently targets the whole world	Geo resolution is fuzzy and scoped to one country. A name unknown in that country returns zero constants, no error — and a campaign with no location criterion targets everywhere. The dry-run cannot catch it: it prints the name you typed	Set geo.country_code per spec. The script aborts the create when any requested name is not covered by a suggestion, and drops suggestions whose search term you never sent or that carry no term at all. It gates on term coverage, never on how many constants came back — one name can yield several constants and another zero, so a count check passes while a zone is silently dropped
Budget off by 10⁶	API takes micros	to_micros(): round(amount * 1_000_000). 9.87 → 9870000
All keywords land BROAD	Syntax is in the string, not a field	Parse [kw]→EXACT, "kw"→PHRASE, else BROAD
Schedule rejected	API only accepts quarter hours	Floor the start, ceil the end, 23:59 → 24:00. Direction matters: ceiling a start opens the day early, flooring an end silently truncates it
Ad rejected after create	Limits discovered at mutate time	Count characters in the dry-run: headline ≤30, description ≤90, path ≤15, 3–15 headlines, 2–4 descriptions
Approved plan ≠ shipped ad	The API caps an RSA at 15 headlines / 4 descriptions / 2 paths, so the create truncates the excess	validate_rsa warns on the maxima, not just the minima: a 20-headline plan prints 20 headlines > max 15 — 5 WILL BE DROPPED on create. A count the human approved must never be silently cut at mutate time — the truncation and the warning read the same constants
Campaign targets the world because a key was missing	An absent geo block is not "no targeting", it is everywhere, and it passes a gate that only fires on names-without-a-country	Worldwide must be typed, never inherited from an omission: no geo.include ⇒ abort, unless the spec says "geo": {"worldwide": true, "note": "<why>"} — and the dry-run then prints *** WORLDWIDE (explicitly requested) ***
Schedule fires at wrong local hour	Windows are interpreted in the account timezone	Express every window in the account timezone. Two distant zones become two overlapping windows in account time, not two timezones
Mutate fails after passing every gate	Account id copied from the UI as 123-456-7890	load_env() strips dashes and rejects a non-numeric id up front

7. Idempotence

python

# name is pre-validated against [A-Za-z0-9_.\-]{1,120}, so it cannot carry a quote into the literal
query = ("SELECT campaign.resource_name FROM campaign "
         f"WHERE campaign.name = '{name}' AND campaign.status != 'REMOVED' LIMIT 1")
# found -> print, return 0, create nothing

The campaign name is the key. Search before create, always. Replaying a run must never produce a second campaign — duplicates compete against each other in the same auction and both spend.

The status filter is not decoration. The campaign report still returns REMOVED campaigns. Without != 'REMOVED', deleting a broken campaign and re-running — the documented recovery move — matches the tombstone, prints "already present", creates nothing, and exits 0. A successful-looking run that did nothing, exactly when you were recovering.

8. Negative keyword taxonomy

Six families that transfer to essentially any B2B account. Ship them at campaign level from day one — they cost nothing and each blocks a class of guaranteed-waste clicks.

Family	Negatives
Price / free	free, cheap, price, prices, pricing, cost, how much, discount, coupon, promo
Jobs / training	job, jobs, hiring, salary, career, internship, training, course, certification, class
Spare parts / DIY	parts, spare parts, battery, replacement, repair, manual, user guide, how to use, diy, refurbished
Marketplaces	amazon, ebay, alibaba, aliexpress, craigslist, marketplace, wholesale directory
Consumer / informational	for home, personal use, definition, what is, wikipedia, forum
Out-of-coverage zones	Every zone you cannot serve, as a keyword and a geo exclusion — belt and braces

These are campaign-level negatives, not a shared list: the script creates one campaign_criterion per negative, attached to this campaign only. Re-created in full for every new campaign. To mutualise them, build a SharedSet once and attach it with CampaignSharedSet — out of scope for this script.

Launch on exact + phrase only. Broad at launch spends the budget discovering that the taxonomy above was right.

9. Tool split

Job	Tool	Gate
Read state, spend, metrics	Read-only API queries or a reporting connector	None — reading is free
Create / edit campaign, ad group, RSA, targeting	SDK script + spec	Human go on the dry-run
Pause / enable, set budget	Vendor UI or a narrow connector	Human, always
Cut at the budget cap	Vendor-native script (§10)	Human installs it, sets the scope, reads the scope: log line, then turns the pause on. After that it is the one unattended action
Offline conversions, reporting	Vendor-native: Data Manager API / UI (UploadClickConversions is closed to new adopters since 2026-06-15), scheduled scripts	Independent of the agent

Hybrid on purpose: write through the API where structure matters, keep caps and conversion ingestion native so they survive the agent being down, refactored, or wrong.

10. The budget cap lives outside the agent

budget-cap-guard.js runs inside the ad account on the vendor's own script runner, hourly:

javascript

var CONFIG = { HARD_CAP: null, WARN_RATIO: 0.9, NAME_CONTAINS: '', PAUSE_ON_CAP: false,
               SHEET_URL: '', ALERTS_TAB: 'budget-cap-guard', RETAIN_DAYS: 400 };
// shipped values THROW: HARD_CAP, NAME_CONTAINS and SHEET_URL are required
// spend >= cap  -> PAUSE_ON_CAP ? pause ENABLED campaigns containing NAME_CONTAINS : alert only
// spend >= 90%  -> append a NEAR_CAP row to its own tab; the agent relays it

Four properties that make it trustworthy:

1. It is not the agent. You do not hand the brake to the driver. If the agent is broken, looping, or hallucinating a budget increase, the cap still fires — vendor scheduler, in the account, no dependency on your runtime.
2. Its only account action is a pause, on a named scope. It cannot raise a budget or enable anything. NAME_CONTAINS is mandatory and validated ([A-Za-z0-9_.-]{3,120}): there is no empty-means-everything mode. Its other writes are rows in the one tab it created, pruned by its own marker.
3. It fails closed. Missing cap, empty prefix, placeholder sheet, "500" as a string, spend that is null or NaN: the run throws before any pause or write. A gate that guesses on bad input is worse than no gate.
4. It holds no secret. No token, no webhook, no credential in a script pasted into a third-party UI.

And two costs, because a cap sold as free is a cap nobody scopes:

• Measured on the whole account, applied to the scope. The cap compares account spend (conservative); the pause hits only the prefixed campaigns. Others keep spending — size HARD_CAP with that in mind.
• No automatic recovery. Spend resets on the 1st; nothing re-enables. Re-activation is deliberately human — so put a calendar reminder next to it, or the failure is silent: "we stopped spending for nine days and nobody noticed."

Set the media budget under the cap (e.g. 450 against a 500 cap) so the buffer absorbs normal variance and the cap only fires on something genuinely wrong. Media is billed by the ad network to the payment method on the account, separate from any agency fee.

11. Editorial red lines

• No prices, no "cheap", no "free" in copy when quoting is a conversation. Those words are negatives — do not bid on what you refuse to say.
• Regulated verticals (health, finance, legal, and anything with a licensing regime) carry vendor policy and statutory constraints. Ad copy and landing page must match; an obligation that applies to one segment must not be implied for a segment it does not cover. The accountable human decides — not the agent, not this skill.

Output format

Post this verbatim after every dry-run. It is the artifact a human approves. It contains no account id — do not add one; the plan is reviewable without it.

text

CAMPAIGN PLAN — [campaign_name]
Spec:      [path]              Segment: [segment]
Budget:    [monthly]/mo -> [daily]/day ([micros] micros)   Share: [share]
Geo:       +[include] / -[exclude] (country [cc])   Schedule: [N] slots, acct tz [tz]
Keywords:  [N] ([N] exact / [N] phrase / [N] broad)   Negatives: [N]
RSA:       [N] headlines / [N] descriptions — limits [OK / N warnings]
Rounding:  [none | HH:MM-HH:MM -> HH:MM/HH:MM]
STATUS ON CREATE: PAUSED (forced in code)
VERDICT: [✅ READY — awaiting human go / ⚠️ N warnings / ❌ blocked: reason]

Filled:

text

CAMPAIGN PLAN — acme-widgets_us_search_demo
Spec:      campaign.example.json     Segment: acme-widgets
Budget:    300/mo -> 9.87/day (9870000 micros)   Share: 100%
Geo:       +United States / -Hawaii (country US)   Schedule: 6 slots, acct tz America/New_York
Keywords:  6 (3 exact / 3 phrase / 0 broad)   Negatives: 44
RSA:       7 headlines / 3 descriptions — limits OK
Rounding:  10:07-23:59 -> 10:00/24:00
STATUS ON CREATE: PAUSED (forced in code)
VERDICT: ✅ READY — awaiting human go

Troubleshooting

Issue	Cause	Fix
"already present -> no create"	Idempotence gate hit on the campaign name	Expected on replay. Rename in the spec, or edit in the UI. If you deleted the campaign and this still fires, your gate is missing the != 'REMOVED' filter (§7)
Second identical campaign appeared	Name changed between runs, gate missed it	Name is the key — keep it stable in the spec
Campaign live everywhere, plan said one country	Geo names did not resolve; no location criterion was created	§6. Set geo.country_code; the create now aborts instead of degrading to worldwide
Campaign exists, no ad, spending	Tree created in several calls, one failed	One mutate (§4). Delete the orphan, re-run — the REMOVED filter makes the re-run actually create

Scope

This skill ONLY: turns a JSON spec into a reviewable plan offline, with no env read and no identifier in the output · creates one campaign tree in one atomic mutate, always PAUSED · searches by name before creating · aborts rather than shipping unresolved or unstated-worldwide geo targeting · reports rounding, micros conversion, and every count or character limit that would change the ad between the plan and the create · ships an opt-in cap that a human installs outside the agent, pauses only a mandatory named scope, fails closed on any invalid input, writes only to a tab it created, and prunes only its own marked rows.

This skill NEVER: enables a campaign · raises a budget · edits live campaigns without a human go · touches display or performance-max inventory · writes copy for a regulated claim without a named human reviewer · hardcodes, prints, or transmits a credential or an account id · bypasses any vendor control, review queue, rate limit, or policy enforcement — if the platform blocks or disapproves something, that is the answer, and a human takes it from there.

---

Alexandre Bloch (Bloch Agents) · ClawHub @AlexBloch-IA