AI Agent Skill (dv360-custom-bidding)

dv360-bidkit ships with an out-of-the-box agent skill: skills/dv360-custom-bidding/SKILL.md.

Equipped with this skill, AI coding assistants (Cursor, Claude Code, Antigravity, GitHub Copilot) act as expert programmatic bidding engineers who never hallucinate unsupported DV360 syntax or uncalibrated weights.

---

Installation into Agent Environments

1. Via skills.sh (Universal Agent Installer)

Install directly with one command:

npx skills add flovv/dv360-bidkit

2. Manual Drop-In

Copy the skill directory into your project's agent folder:

# For Cursor / Antigravity:
mkdir -p .agents/skills/dv360-custom-bidding
cp -r skills/dv360-custom-bidding/* .agents/skills/dv360-custom-bidding/

# For Claude Code:
mkdir -p .claude/skills/dv360-custom-bidding
cp -r skills/dv360-custom-bidding/* .claude/skills/dv360-custom-bidding/

---

Agent Guidelines Enforced by the Skill

When the skill is active, AI models are constrained by strict domain rules:

1. The Engine Executes All Calculations:
   The AI assistant will never write custom heuristic formulas or ad-hoc viewability calculations in Python. It always invokes dv360_bidkit functions (lint_script(), compile_script(), evaluate_script(), fit_display_quality_proxy(), or evaluate_and_calibrate_funnel_weights()).
2. Model Training Exclusions vs. Ad-Serving Blocks:
   In DV360, return None excludes an impression from ML model training; it does NOT block the ad from serving. The agent will never claim custom bidding scripts enforce publisher blacklists. Ad-serving blocks must be exported to line-item targeting exclusion lists (targeting_block_request.csv).
3. Econometric Transition Caps on Proxies:
   The agent will enforce $w_{\text{proxy}}\le w_{\text{macro}}\times P(\text{Macro}\mid \text{Proxy})$, preventing uncalibrated proxy values from dominating model training.
4. Outlier Capping on Continuous Distributions:
   The agent automatically enforces ternary caps ((cap if val > cap else val)) on revenue and duration signals at the 95th percentile to prevent pacing controller destabilization.
5. No Runtime Wildcard Subdomain Claims:
   The agent knows Google's scoring engine evaluates exact string equality and numeric ranges, and will not emit regex or dynamic .endswith() patterns in compiled scripts.

---

Example Agent Prompt

Once the skill is installed, you can simply ask your AI coding assistant:

"Analyze client_q3_report.csv, map the columns, check whether we have enough historical volume for Google S2 readiness, and compile a 3-tier viewability and Floodlight conversion bidding script."

The agent will load the report using ReportMapping, simulate performance with evaluate_script(), diagnose readiness metrics, and output verified, production-ready DV360 DSL.