Run advertising campaigns like a seasoned programmatic trader by talking to your AI assistant in plain language. This is an open library that teaches Codex and other skill-aware agent tools the real decision rules, checklists, and steps an experienced trader, analyst, and ad-operations specialist uses, across the major advertising platforms and the measurement and tagging stack underneath them (Google Analytics 4, Campaign Manager 360, and Google Tag Manager), so the buy and the signal that judges it live in one place.
Programmatic advertising means buying digital ads automatically through live auctions, across platforms like Google's Display & Video 360, Google Ads, Amazon DSP, StackAdapt, and The Trade Desk. Each platform has its own rules, its own jargon, and its own ways to go wrong, and the knowledge to run them well lives in the heads of a few experienced people.
If you ask a general AI assistant a real trading question ("why is my campaign not spending?", "which bid strategy should I use?", "build me a plan for a fifty thousand dollar video campaign"), it tends to give generic or slightly wrong answers, because it does not know the specifics of each platform.
This library fixes that. It gives the assistant the expert's playbook, so its answers match what a senior trader would actually do.
Think of it as three layers.
- Skills are the knowledge. Each skill is a short written playbook for one job: how to structure a campaign, how to choose a bid strategy, how to read a report, how to find why a campaign is not delivering. When your question matches a skill, the assistant reads that playbook and follows it. There are dozens of them, grouped by platform.
- Agents are the roles. Fourteen ready-made specialists that use the skills: a media planner, a trader who builds the campaign, an optimizer who improves it, an account-operations specialist, a measurement-and-tagging engineer who owns the conversion signal, an ad-ops trafficker, a paid-search and PMax specialist, a data-and-analytics engineer who runs the warehouse, a measurement analyst who reconciles the numbers, a privacy-compliance officer who gates the data, a programmatic ops lead who runs the queue, a reporting analyst, a client-communications lead who writes the client update, and a scrutinizer who double-checks the work before it goes out. They check each other and pass work back on failure rather than relaying it forward. See docs/AGENT-FLOWS.md.
- Loops are the routines. Repeatable checklists you can run on demand or on a schedule, like a daily pacing check or a weekly optimization pass.
There are also simple calculators that need no setup, and a guide for connecting the assistant to your live accounts when you want it to read your real data.
You do not configure anything to start. You install it, then ask your assistant questions in normal language, and it picks the right skill on its own.
- It is based on real documentation, not guesses. Almost every claim links to the platform's own official help page, and the few platforms with private documentation (notably The Trade Desk) are clearly marked as such.
- It is safe by design. The assistant reads, analyzes, and recommends. It never spends your money or changes a live campaign on its own. A person always approves any change. This matches what the ad platforms themselves allow today.
- It checks its own work. A built-in scrutinizer reviews reports and recommendations for math errors, overclaiming, and unsupported statements before they reach a client.
- One way of working across every platform. The same agents and routines run on any of the platforms; only the platform-specific knowledge changes underneath.
- It was reviewed by a panel of trader and ops experts and their feedback was built in. See docs/CRITIQUE-AND-ROADMAP.md.
A fair question: do the platforms already have this? Yes. Every major DSP now ships its own AI agent. The Trade Desk has Koa Agents, Google has Ask Advisor, its Gemini agent that now spans DV360 and Google Ads, Amazon has Ads Agent, and StackAdapt has Ivy, which even has a feature called "Skills" and a connector for external AI workflows.
They are good. On their own platform they do things this package does not: they have live access to your account, they build and optimize campaigns in product, they are deeper, and they are free in the interface. This package does not try to beat them there. It does the one thing none of them does, and structurally none of them will:
| A platform's native agent | This package | |
|---|---|---|
| Scope | One platform, locked in | All five in one assistant, with shared foundations |
| Whose side it is on | The seller's. It will not move budget off its own platform | The buyer's. It can recommend shifting spend between platforms or away from one |
| Can it act | Yes, it builds and optimizes live in product | It advises; a person executes |
| Depth on its platform | Deepest, real time, always current | A researched public snapshot |
| Openness | Closed | Open text you can read, audit, and fork |
| Where it runs | The platform's walled garden, on its model | Your own assistant, any model, including a local one |
The point is "and," not "versus." A native agent is the best tool for acting inside its own platform. This package is the vendor-neutral layer above all of them: one consistent way to plan and reason across five DSPs, on the buyer's side, in your own assistant. It can even orchestrate the platforms' own connectors (Google Ads, Amazon Ads, and StackAdapt each expose one) rather than replace them. No single platform will build a neutral agent that helps you run a competitor's DSP or move money off its own inventory. That gap is what this fills.
A few real examples of the difference it makes.
"My DV360 campaign is not spending its budget. What is wrong?" On its own, an assistant guesses. With this library, it runs the troubleshooting playbook in the right order (status, then budget and flight, then bid and win rate, then how narrow the targeting is, then inventory, then the creative) and tells you the single binding cause and the fix, the way a senior trader debugs it.
"Build me a media plan for a fifty thousand dollar connected-TV awareness campaign." The planner agent produces a structured plan with the right success metric (reach and frequency, not clicks), the right platform for CTV, a budget split, a flight, and a measurement plan, then hands it to the scrutinizer to check before it reaches the client.
"Compare these two buys for me." The eCPM calculator converts a cost-per-click buy and a cost-per-action buy to the same effective cost per thousand impressions, so you can compare lines that were priced differently.
The panels below show how the pieces fit together, how specialist agents hand work off, how
the assistant applies a skill, and how the offline tools turn routine checks into gates. PNG
exports are used here for reliable previews, with matching editable SVG files in images/.
Skills hold the knowledge, agents are the specialist roles that use them, and loops are the repeatable routines. Underneath it all, the assistant only recommends. A person approves any change.
The roles create a production workflow, not a chain of generic prompts. The scrutinizer can send work back to the owner with concrete fixes before it reaches a client or a live account.
Ask why a campaign is not spending, and instead of a generic answer the assistant walks the same triage a senior trader uses (status, budget, bid, targeting, inventory, creative), then names the one binding cause, here a bid below the floor, and the fix. It stops at a recommendation and does not touch the account.
The simple calculators need no accounts and no setup. Give the budget and flight planner your budget, your flight, and a target CPM, and it returns the daily budget, the impressions to expect, and pacing checkpoints to catch drift early.
Programmatic traders, ad-operations specialists, and analysts who want an assistant that already knows the platforms, agencies that want consistent quality, and anyone learning programmatic who wants the decision rules an experienced trader applies.
A library of agent skills organized by platform, plus shared foundations and reporting, the fourteen specialist agents, and the loop library. Covered today: the demand-side platforms (DV360, Google Ads, Amazon DSP, StackAdapt, The Trade Desk), the Google Marketing Platform measurement and tagging stack (Google Analytics 4, Campaign Manager 360, Google Tag Manager, Search Ads 360), server-side conversions and campaign support for Meta, TikTok, and Microsoft Advertising, retail media and curation, creative testing, programmatic operations, and a privacy and compliance layer.
| Skill | What it does |
|---|---|
programmatic-foundations |
Glossary, auction and KPI math, funnel model, and the trader, analyst, and ops mental model every platform skill builds on. |
reporting-by-campaign-goal |
State-of-the-art report recipes per objective: awareness, consideration, conversion, retention, and reach planning. |
path-to-conversion-analysis |
Multi-touch paths: touchpoints to convert, time lag, top paths, and assisted conversions, via CM360, GA4, and Ads Data Hub. |
dsp-selection |
Which demand-side platform to use for which goal, and the tradeoffs that decide it. |
reach-and-frequency-planning |
Deduplicated reach across platforms, effective frequency, the reach curve, and the identity limits. |
incrementality-and-experimentation |
Lift testing done right: conversion lift, geo lift, holdouts, power and sample size, reading a result. |
direct-response-creative-testing |
Creative test design, asset-level reads, fatigue signals, holdout needs, and what to brief next. |
client-deliverable-templates |
Fill-in media plan, QBR deck, proposal, plain-English glossary, and bad-news framing for clients. |
value-based-bidding |
Feeding accurate conversion values (revenue, margin, LTV, new-customer value) so automated bidding optimizes to profit. |
bid-landscape-and-win-rate |
Reading the win-rate-by-bid curve to find the efficient bid, marginal CPA versus volume, first-price effects. |
marketing-mix-modeling |
What MMM is and when to use it, data needs, Meridian and Robyn, reading contribution and response curves. |
data-quality-and-reconciliation |
Why conversion numbers differ across tools, acceptable bands, pre-ship checks, and a real anomaly method. |
discrepancy-and-reconciliation |
Ad server versus DSP impression discrepancies, tolerance bands, make-goods, and a month-end close. |
tag-and-pixel-governance |
Floodlight and pixel setup and validation, Consent Mode, deduplication, and a pixel inventory and retirement policy. |
change-management-and-incident-response |
Maker-checker approvals, a bulk-edit pre-flight, and an incident runbook with severity tiers. |
partner-and-advertiser-onboarding |
A gated sequence from signed insertion order to first-campaign-ready, with billing and measurement checks. |
brand-safety-and-suitability |
Pre-bid versus post-bid, MFA and invalid traffic, suitability tiers, regulated categories, supply-path transparency. |
privacy-and-consent |
GDPR, CCPA and CPRA, Consent Mode, TCF, identity consent, and the cookieless and Privacy Sandbox state. |
trader-onboarding |
A week 1, 2, and 4 ramp through the library for a new trader, ending in a graded build. |
approval-and-escalation-governance |
Who approves what, escalation paths, service levels, and a trader capability model. |
cross-platform-conversion-reconciliation |
One deduplicated conversion count across GA4, CM360, and each DSP, with the overlap stripped and a stated method. |
gmp-integration-floodlight-ga4-linking |
How GA4, Campaign Manager 360 Floodlight, DV360, and Search Ads 360 link and share audiences and conversions across the Google Marketing Platform. |
marketing-analytics-warehouse-and-dashboards |
A conformed cross-platform reporting model and dashboard-ready marts built on one set of definitions. |
pipeline-orchestration-and-data-quality |
Idempotent load pipelines, watermarks, freshness SLAs, and schema-drift detection so a rerun never double counts and a broken feed never ships. |
ctv-measurement-currencies |
The post-Nielsen currency landscape for CTV and video, and how to plan and reconcile across measurement providers. |
| Skill | What it does |
|---|---|
programmatic-ops-control-center |
Daily command center for launches, live delivery, measurement, creative, tickets, support, access, incidents, and close risk. |
ops-ticket-intake-and-triage |
Turns messy requests into actionable tickets with required fields, severity, SLA, owner, evidence, and acceptance criteria. |
cross-platform-launch-qa |
Platform-neutral go/no-go gate for campaign launches across DSPs, paid social, search, ad server, tracking, consent, and reporting. |
bulk-change-governance-and-rollback |
Scope, snapshot, diff, canary, approval, read-back verification, and rollback package for large account changes. |
platform-support-escalation |
Evidence-based support cases for platform, publisher, API, creative, deal, tracking, reporting, access, and billing issues. |
cross-platform-taxonomy-governance |
Campaign, creative, UTM, cost-center, and reporting naming that joins cleanly across operations, analytics, and finance. |
platform-access-and-permissions-governance |
Least-privilege access intake, role assignment, offboarding, quarterly review, and break-glass control. |
creative-review-and-approval-queue |
Creative intake, spec QA, brand and legal approval, platform review, trafficking readiness, rejections, swaps, and live verification. |
programmatic-month-end-close |
Delivery, spend, invoice, discrepancy, make-good, accrual, taxonomy, and finance handoff close process. |
| Skill | Job | What it does |
|---|---|---|
dv360-campaign-architecture |
Trading | Partner to advertiser to campaign to insertion order to line item structure, and when to split. |
dv360-bid-strategy |
Trading | Fixed, automated, and custom bidding. Target CPA, CPM, ROAS. Learning periods and pitfalls. |
dv360-targeting-and-audiences |
Trading | First-party and Google audiences, combination logic, geo, device, contextual, viewability and IVT. |
dv360-deals-and-inventory |
Trading | Open auction, PMP, Programmatic Guaranteed, Preferred Deals. Activation and non-delivery fixes. |
dv360-frequency-and-brand-safety |
Trading | Frequency caps, content and publisher exclusions, DoubleVerify and IAS, viewability standards. |
dv360-pacing-and-optimization |
Trading | Pacing modes, pacing math, under and over-delivery fix trees, impression loss diagnosis. |
dv360-youtube-and-video |
Trading | YouTube and video line items: skippable, non-skippable, bumper, in-feed, Shorts, CPV, and video reach campaigns. |
dv360-creative-trafficking |
Ops | Third-party tags, VAST, click macros, secure tags, and a creative QA checklist for what blocks a creative from serving. |
dv360-video-creative-specs |
Ops | Hosted video, third-party VAST, companions, transcodes, tracking URLs, duration, aspect ratio, and CTV spec checks. |
dv360-reporting |
Analytics | Offline vs instant reporting, report types, the metric and dimension glossary, scheduling. |
dv360-measurement-and-attribution |
Analytics | Floodlight, Campaign Manager 360, attribution models, Brand Lift, reach and frequency. |
dv360-advanced-analytics-adh |
Analytics | Ads Data Hub, privacy checks, BigQuery Data Transfer, joining first-party data. |
dv360-custom-bidding |
Analytics | Rule-based, script, and Ads Data Hub custom bidding. Scoring, attribution, staged rollout. |
dv360-account-setup-and-taxonomy |
Ops | Partner and advertiser setup, naming conventions, roles and permissions, governance. |
dv360-launch-qa |
Ops | Pre-flight QA checklist and sign-off workflow before any campaign goes live. |
dv360-troubleshooting |
Ops | Ordered playbooks for no delivery, pacing, win rate, viewability, creatives, conversions. |
dv360-api-and-sdf-automation |
Ops | DV360 API v4 resources, Structured Data Files v10, and a safe-to-automate matrix. |
ctv-supply-and-ad-pods |
Trading | Connected TV supply, ad pods and pod deduplication, PG and curated CTV deals, and frequency across apps and devices. |
| Skill | Job | What it does |
|---|---|---|
google-ads-account-structure |
Structure | Account and manager (MCC) hierarchy, campaign and ad group organization, the shared library, and limits. |
google-ads-campaign-types |
Structure | Search, Performance Max, Demand Gen, Display, Video, Shopping, and App, with an objective-to-type guide. |
google-ads-performance-max |
Campaigns | Asset groups, audience signals, listing groups, search themes, brand exclusions, and PMax versus Search. |
google-ads-bidding |
Bidding | Smart Bidding (tCPA, tROAS, maximize conversions or value), manual CPC, portfolio strategies, bid adjustments. |
google-ads-keywords-and-match-types |
Search | Broad, phrase, and exact match, negatives, the search terms report, and keyword research. |
google-ads-audiences-and-targeting |
Targeting | Audience segments, Customer Match, targeting versus observation, optimized targeting, content targeting. |
google-ads-budgets-and-pacing |
Budget | Average daily budgets, the 2x daily and monthly cap behavior, shared budgets, and limited-by-budget. |
google-ads-conversion-tracking-and-attribution |
Measurement | Conversion actions, Enhanced Conversions, Consent Mode, primary vs secondary, and attribution models. |
google-ads-reporting |
Analytics | The report editor, custom columns, segments, the impression-share metrics, scripts, and GAQL. |
google-ads-optimization-and-troubleshooting |
Ops | Optimization score, learning and limited statuses, disapprovals, low impression share, delivery fixes. |
google-ads-api-and-bulk-operations |
Automation | Google Ads API v24, GAQL, Editor, scripts, and a safe-to-automate matrix with a report puller. |
google-ads-shopping-and-feed |
Retail | Merchant Center, product feed quality and attributes, disapprovals, and Shopping versus Performance Max. |
| Skill | Job | What it does |
|---|---|---|
amazon-dsp-account-structure |
Structure | Advertiser, order, and line item hierarchy, managed vs self-service, product types, and the Amazon Ads pixel. |
amazon-dsp-campaign-setup |
Campaigns | Building orders and line items: supply, budget, pacing, flight, goal, frequency, dayparting, and targeting. |
amazon-dsp-audiences |
Targeting | Amazon shopping and streaming audiences, advertiser and AMC audiences, lookalikes, and ASIN retargeting. |
amazon-dsp-inventory-and-supply |
Supply | Amazon owned-and-operated (Prime Video, Fire TV, Twitch, IMDb), deals, and third-party exchanges. |
amazon-dsp-bidding-and-optimization |
Bidding | Optimization goals (reach, CPA, ROAS, VCR, DPVR), bid, supply, and audience optimization, and pacing. |
amazon-dsp-creative-and-formats |
Creative | Display, online video, streaming TV, audio, and responsive e-commerce creatives, and where to get specs. |
amazon-dsp-measurement-and-reporting |
Analytics | The retail funnel: detail page views, purchases, ROAS, new-to-brand, reach, frequency, and attribution. |
amazon-marketing-cloud |
Analytics | The AMC clean room: SQL on event-level signals, custom attribution, overlap, incrementality, and audiences. |
amazon-dsp-api-and-automation |
Automation | The Amazon Ads API for DSP, reporting and audiences APIs, the AMC API, access gating, and safe-to-automate. |
amazon-ads-agentic-and-mcp |
Automation | Amazon's agentic-AI surface and MCP connector: what it automates, the access model, and where a human stays in the loop. |
| Skill | Job | What it does |
|---|---|---|
stackadapt-account-structure |
Structure | Account, campaign, ad group, and ad hierarchy across native, display, video, CTV, audio, and DOOH, and the pixel. |
stackadapt-campaign-setup |
Campaigns | Building a campaign and ad groups: channel, objective, budget, flight, pacing, bid, targeting, and creatives. |
stackadapt-targeting-and-audiences |
Targeting | Retargeting, lookalikes, third-party data, custom segments, and StackAdapt's contextual targeting. |
stackadapt-bidding-and-budgets |
Bidding | Automated and manual bidding, goal types, maximum bids, budget setting, and pacing. |
stackadapt-creative-and-formats |
Creative | Native, display, HTML5, video, CTV, audio, DOOH, Creative Studio, creative QA, and review blockers. |
stackadapt-inventory-and-brand-safety |
Supply | Exchange and PMP supply, CTV inventory, exclusion lists, contextual avoidance, and verification. |
stackadapt-reporting-and-attribution |
Analytics | Reporting dashboards and exports, the pixel and event tracking, UTMs, and the attribution approach. |
stackadapt-optimization-and-troubleshooting |
Ops | Triage and symptom playbooks for delivery, pacing, performance, creatives, and conversion tracking. |
stackadapt-api-and-automation |
Automation | The StackAdapt GraphQL API (request-only access), reporting, and a safe-to-automate matrix. |
The Trade Desk's operational knowledge base and API reference sit behind a partner login, so these skills are written at the public-concept level from TTD's public pages and the open Unified ID 2.0 documentation. They state the model and flag where exact menus, fields, and numbers must be confirmed in the partner platform, rather than inventing specifics.
| Skill | Job | What it does |
|---|---|---|
ttd-platform-overview |
Overview | What The Trade Desk is, the independent open-internet DSP, Kokai, Koa AI, channels, and routing. |
ttd-campaign-structure |
Structure | The account and campaign hierarchy and where settings live, at the public concept level. |
ttd-targeting-and-audiences |
Targeting | First and third-party data, the data marketplace, contextual, and seeds, with specifics flagged. |
ttd-bidding-and-optimization |
Bidding | Koa AI valuation, seeds and bid factors, predictive clearing, Performance mode, and forecasting. |
ttd-inventory-and-deals |
Supply | Open market, private marketplace, Programmatic Guaranteed, and the OpenPath supply path. |
ttd-creative-and-formats |
Creative | Display, HTML5, video, CTV, audio, DOOH, native, companions, VAST, and publisher spec QA from public TTD specs. |
ttd-identity-and-uid2 |
Identity | Unified ID 2.0 and EUID: tokens, operators, refresh, integration paths, and OpenPass. |
ttd-measurement-and-reporting |
Analytics | Reporting and attribution concepts, with the gated platform specifics flagged. |
ttd-api-and-automation |
Automation | The partner-gated TTD API model and a safe-to-automate posture. |
ttd-ctv-and-video-buying |
Video | CTV and online-video buying on The Trade Desk: inventory, deals, forecasting, and completion, at the public-concept level. |
The measurement backbone. GA4 defines the events and key events that every buying platform imports as conversions, and the audiences it activates. GA4 uses "key events", not the old "conversions" label.
| Skill | Job | What it does |
|---|---|---|
ga4-property-and-data-stream-setup |
Setup | Property and data-stream setup, the Google tag or GTM install, enhanced measurement, and data controls. |
ga4-events-and-key-events |
Setup | The event data model, reserved names and limits, and marking and managing key events and their Ads counterparts. |
ga4-conversions-and-audiences |
Activation | Building audiences and exporting conversions and audiences to Google Ads, DV360, and SA360, respecting the 2026 consent change. |
ga4-explorations-and-reporting |
Analytics | Free-form, funnel, path, cohort, and overlap explorations, and reading standard reports within sampling and cardinality limits. |
ga4-bigquery-export-and-sql |
Analytics | The daily and streaming BigQuery export, its nested event schema, and unsampled event-level SQL. |
ga4-apis-and-measurement-protocol |
Automation | The Data API, Admin API, and Measurement Protocol for reporting, configuration, and server-side events. |
The ad server and the conversion source DV360 bids on. Floodlight is where conversions are defined, so this is the layer a trader optimizing to conversions depends on.
| Skill | Job | What it does |
|---|---|---|
cm360-trafficking-and-ad-tags |
Ops | Advertiser to campaign to placement to ad to creative structure, ad tags, click commands, and VAST wrapping. |
cm360-floodlight-and-conversions |
Measurement | Floodlight activities and groups, counting methods, variables, and aligning definitions with the buy side. |
cm360-reporting-and-trafficking-api |
Automation | Report Builder, the Trafficking API for programmatic build and verification, and pulling a trafficking sheet. |
cm360-data-transfer-and-attribution |
Analytics | Data Transfer v2 event files, attribution models, and joining delivery to conversions at the event grain. |
cm360-verification-and-active-view |
Analytics | Active View viewability, verification and IVT signals, and how they reconcile with the DSP and a third party. |
The tag layer, client-side and server-side, and where Consent Mode lives.
| Skill | Job | What it does |
|---|---|---|
gtm-web-container-and-datalayer |
Setup | The web container, a clean dataLayer contract, and the naming that keeps tags and triggers reliable. |
gtm-tags-triggers-variables |
Setup | Tags, triggers, and variables for GA4, Ads, and Floodlight, and keeping the container aligned to the measurement plan. |
gtm-consent-mode-v2 |
Consent | Basic versus advanced Consent Mode v2, the four consent signals, and wiring tags to respect granted and denied states. |
gtm-server-side-tagging |
Setup | The tagging server, the GA4 client, transport, and moving collection server-side for durability and control. |
gtm-server-conversion-enhancement |
Measurement | Enhanced conversions and server-side conversion APIs that recover signal the browser drops, without double counting. |
gtm-qa-versioning-and-api |
Ops | Workspaces, versioning, Preview, a QA checklist, and the GTM API for programmatic container management. |
Cross-engine search management next to programmatic.
| Skill | Job | What it does |
|---|---|---|
sa360-account-structure-and-engines |
Structure | Engine accounts across Google, Microsoft, and others, mirrored structure, labels, and one taxonomy. |
sa360-bidding-and-budget-management |
Bidding | SA360 bid strategies, budget management and portfolios, and pacing that neither starves nor overspends. |
sa360-conversions-and-reporting-api |
Measurement | Conversion import and validation, Floodlight and GA4 links, and the SA360 reporting API. |
Server-side conversions first, because that is where signal and match quality are won.
| Skill | Job | What it does |
|---|---|---|
meta-conversions-api-and-datasets |
Measurement | The Conversions API and datasets, hashing and match quality (EMQ), deduplication with the pixel, and the access model. |
meta-campaign-setup-and-optimization |
Campaigns | Meta campaign objectives, Advantage+ versus manual structure, audiences, placements, budgets, learning, creative, and reporting. |
meta-advantage-plus-campaigns |
Campaigns | Advantage+ Sales and audience automation, the levers that still matter, and where the model needs steering. |
tiktok-events-api-server-side |
Measurement | The TikTok Events API, server-side event delivery, deduplication, and identity signals. |
tiktok-creative-and-catalog-optimization |
Campaigns | TikTok creative, Smart+ inputs, catalog health, product sets, creator-style ads, fatigue, and campaign optimization. |
tiktok-smart-plus-and-campaigns |
Campaigns | Smart+ automation and campaign structure, and the inputs that decide whether it performs. |
microsoft-uet-and-conversions-api |
Measurement | UET tags, the Conversions API, offline conversions, and consent handling on Microsoft Advertising. |
microsoft-campaign-setup-and-optimization |
Campaigns | Microsoft Search, PMax, Shopping, Audience, imports, feeds, budgets, brand controls, and optimization QA. |
microsoft-performance-max-and-shopping |
Campaigns | Microsoft Performance Max and Shopping, feed health, and where it mirrors or diverges from Google. |
| Skill | Job | What it does |
|---|---|---|
retailmedia-foundations-and-networks |
Retail | Retail media networks beyond Amazon (Walmart Connect and others): onsite, offsite, and the commerce-data model. |
retailmedia-onsite-sponsored-products |
Retail | Onsite sponsored products and their auction, bidding, and reporting mechanics across networks. |
curation-and-curated-marketplaces |
Supply | Curated marketplaces and single curated Deal IDs that carry audience and inventory, and how to buy them. |
curation-and-supply-path-optimization |
Supply | Supply-path optimization: choosing paths, deduping sellers, and cutting hops without cutting scale. |
supply-chain-transparency-adstxt-sellersjson |
Supply | ads.txt, sellers.json, and the SupplyChain object, and how to read them to verify a legitimate path. |
| Skill | Job | What it does |
|---|---|---|
privacy-sandbox-and-signal-loss-2026 |
Privacy | The 2026 ground truth (Privacy Sandbox ad APIs retired, cookies stayed), what survives, and durable alternatives. |
consent-signal-verification-and-decode |
Privacy | Decoding a TCF or GPP string and confirming it carries the purposes and vendors an activation depends on. |
data-subject-rights-and-subprocessor-governance |
Compliance | DSAR and opt-out fulfillment across platforms, the sub-processor and DPA register, and an append-only evidence ledger. |
The package also ships specialist agents that use the skills above to run a full workflow. Installed with the plugin they are auto-discovered, and you can call one by name.
| Agent | Role |
|---|---|
media-planner |
Turns a brief into a media plan: objective, KPI, audience, inventory, budget, and measurement. |
programmatic-trader |
Builds the campaign from the plan: structure, bidding, targeting, deals, frequency, pacing. |
optimization-specialist |
Optimizes and troubleshoots in flight, one lever at a time, with expected impact. |
account-operations-specialist |
Sets up the account, enforces taxonomy, runs launch QA, and handles safe bulk operations. |
programmatic-ops-lead |
Runs the operations desk: daily queue, launch gate, tickets, bulk-change safety, support escalation, creative, access, and close. |
measurement-and-tagging-engineer |
Owns the conversion signal end to end: GA4 events and key events, Floodlight, GTM, consent, and server-side. |
ad-ops-trafficker |
Traffics the ad server (CM360 placements, ads, creatives, tags) and QAs every tag before it reaches the trader. |
paid-search-and-pmax-specialist |
Runs Google Ads Search, Shopping, and Performance Max, and SA360 across engines. |
data-and-analytics-engineer |
Builds and guards the reporting warehouse: BigQuery models, exports, pipelines, freshness, and schema drift. |
measurement-analyst |
The neutral cross-platform read: reconciles totals, strips double counting, separates assist from lift. |
privacy-compliance-officer |
Gates data activation by region, verifies consent, and runs DSAR and sub-processor governance. |
reporting-analyst |
Builds the right report per goal and runs measurement, attribution, and path to conversion. |
client-communications-lead |
Translates results into clear, honest, client-ready communication. |
qa-scrutinizer |
Independent reviewer that scores and gates builds, reports, and client comms before they ship. |
The agents form a closed loop, not a one-way relay. Three of them are reviewers with the
authority to hold work and pass it back to its author: qa-scrutinizer gates every build,
report, and message; measurement-analyst independently checks the numbers; and
privacy-compliance-officer clears or holds any data activation. A typical flow:
media-planner, then programmatic-ops-lead, account-operations-specialist, measurement-and-tagging-engineer, and
ad-ops-trafficker to set up and instrument, then programmatic-trader (or
paid-search-and-pmax-specialist) to build, then optimization-specialist in flight, then
data-and-analytics-engineer, measurement-analyst, reporting-analyst, and
client-communications-lead to measure and report, with the three reviewers gating each
handoff and returning anything that fails. The full map is in
docs/AGENT-FLOWS.md.
The loops/ folder is a catalog of repeatable agent loops for daily and weekly trading,
measurement, and reporting work: a pacing sweep, an optimization pass, a pre-launch QA gate,
budget reallocation, creative fatigue, anomaly detection, search-term mining, brand-safety
monitoring, client reporting, and business-review prep, plus a measurement and compliance set,
a conversion-tracking health check, a consent-mode compliance sweep, a GA4 data-quality
monitor, a cross-platform conversion reconciliation, a Floodlight and tag pre-launch gate, a
server-side container monitor, an identity and deprecation watch, a search and PMax query
sweep, a pipeline freshness and schema-drift watch, an ops command-center daily sweep, a
bulk-change safety gate, and a month-end close loop. Each loop is a bounded feedback cycle
with an observable success gate and a named stopping condition. Every loop monitors and
recommends rather than spending on its own, and any change is gated on human approval. See
loops/README.md for the catalog and how to run a loop on demand or on a
schedule.
The templates/ops/ folder includes copy-ready artifacts for ticket intake, platform
escalation, bulk-change plans, cross-platform launch QA, taxonomy, access review, creative
review, month-end close, and the daily ops command center. They are plain Markdown or CSV so
they can move into Jira, Asana, Sheets, docs, or a client-approved storage location.
WORKFLOWS.md shows how the three layers compose into end-to-end workflows that run the same way on every platform: launch a campaign, run it in flight, and report to the client, with the QA scrutinizer gating each handoff. To move to a different demand-side platform, the agents and loops stay the same and only the platform skill set changes. That is what makes this a multi-platform operating system rather than five separate playbooks.
Add the repository as a plugin source. Codex reads .codex-plugin/plugin.json, which points
at the shared skills/ directory.
Clone the repo and run the installer. It symlinks each skill into the runtime skills
directories so a git pull keeps them current.
git clone https://github.com/scumunna/programmatic-skills.git
cd programmatic-skills
./install.sh # symlink skills and agents into the runtime directories
./install.sh --copy # copy instead of symlink
~/.agents/skills is the shared path read by Codex, Copilot CLI, and Gemini CLI, so a single
install covers all of them. Agents are also symlinked into ~/.codex/agents.
The skills are plain markdown, so the knowledge works with any model, not just hosted ones.
Skill-aware harnesses such as Codex load them automatically. For a model with no skill
harness, including a local model through Ollama or LM Studio or any OpenAI-compatible endpoint,
tools/skill_router.py routes a question to the right skill and prints a prompt or calls the
model directly. See docs/USING-ANY-LLM.md.
Talk to your agent normally. Skills activate when your request matches what a skill covers, for example "structure a DV360 prospecting campaign for three markets" or "my line item is underpacing, what do I check". Each skill carries the decision rules, checklists, and templates the agent needs to respond like a practitioner.
The tools/ folder has two kinds of helpers. A set of no-setup calculators and validators
(compare buys on a common eCPM, check whether a frequency cap can deliver, plan a budget across
a flight, audit Floodlight activities, QA a GA4 event export, decode a TCF or GPP consent
string, reconcile conversions across platforms, lint UTM tags, validate that expected tags
fired, check warehouse freshness, and locate a data subject's records for a DSAR) that run on
any machine with Python 3, and read-only report pullers bundled with the platform skills that
read from your own account. Everything is read-only and reads credentials from the environment,
never hardcoded. See tools/README.md.
Everything here reads and recommends; nothing changes a live campaign on its own. To give an agent live platform access safely, including how to connect the existing official MCP servers and why any spend-affecting change stays behind a human, see docs/CONNECTING-TOOLS.md.
Each platform has an honest guide for getting your real campaign data into the assistant, from the no-setup path (export a report and hand it over, which works everywhere with any model) to the API and MCP paths, with each platform's real access gating spelled out:
- Google Ads: docs/DEMO-GOOGLE-ADS.md
- DV360: docs/CONNECT-DV360.md
- Amazon DSP: docs/CONNECT-AMAZON-DSP.md
- StackAdapt: docs/CONNECT-STACKADAPT.md
- The Trade Desk: docs/CONNECT-TTD.md
The structure is built for more platforms. Cross-platform concepts live in
programmatic-foundations. A new platform is added as its own prefixed set of skills, for
example ttd-* for The Trade Desk or ga4-* for Google Analytics 4, without restructuring.
The library now spans the demand side, the Google Marketing Platform measurement and tagging
stack (GA4, CM360, GTM, SA360), server-side conversions and campaign operations for Meta,
TikTok, and Microsoft, retail media and curation, creative testing, and privacy and
compliance. The roadmap skills from the practitioner critique are now shipped, and v1.3 adds
the programmatic operations layer that working desks need to run queues, launch gates, bulk
changes, support escalations, access, creative review, taxonomy, and month-end close. What
remains is real-world validation by working traders and deliberately human-gated live account execution. See
docs/CRITIQUE-AND-ROADMAP.md and CONTRIBUTING.md for the recipe.
python3 scripts/validate_skills.py
This checks every skill and agent: frontmatter, naming, description length, the no-em-dash writing standard, that reference links resolve, and that each agent only references skills that exist.
This is an independent, unofficial project. It is not affiliated with, endorsed by, or sponsored by Google. Display & Video 360, DV360, Campaign Manager 360, and Ads Data Hub are trademarks of Google LLC. Platform behavior changes over time. Verify against the current official documentation before acting on a live account.
MIT. See LICENSE.