arbo-skills

Find test adsets that have proven themselves and clone them into cost cap adsets, quickly and safely. It studies how past cost cap clones turned out to calibrate the thresholds (profit, ROI, conversions, RPC) and the cap, then scans today's active uncapped adsets, applies the clone rules (last cap clone, cap adsets of the offer per account and per country), suggests a cap per adset from its RPC, and submits the approved ones with submit_adset_clones_with_cap / submit_adset_clones_with_cap_and_budget. Use when the user asks to find cost cap candidates or winners to scale, to clone winners into cost cap, to "promote" test adsets, to check which adsets are ready for a cap, or to review how earlier cap clones did.

Design a new base prompt for CK's AI ad-image generation (the template that turns an article headline and description into a Facebook image ad), from the user's description of the desired look and/or reference images. Produces a Facebook-compliant, CTR-oriented, placeholder-correct prompt with a short camelCase name, a description and variations. Use when the user asks to create, design, rework, review or vary an image prompt, image template, base prompt, ad image style or creative style, or sends reference ad images and wants a prompt that makes images like them. Also use when the user wants a test prompt (an existing base prompt filled with a given title for manual testing), or wants to list, read, submit or edit saved image prompts through the CK MCP (list_image_prompts, get_image_prompt, save_image_prompt).

Review and rebalance a division's automated CK arbo scripts as one system. Finds scripts that work against each other (one scales or activates what another cuts, pauses or kills), checks whether the priority order lets the important scripts act first, spots scripts that break another script's intent and need a rewrite, and finds scripts that currently have no effect (never run, never match, or lose all their adsets to higher-priority scripts). Simulates a full automation pass with dry runs, reads the tag history for real conflicts, then proposes a new priority order and rewrites, and applies them only when the user approves (update_stat_script_code). Use when the user asks to review, audit, check, rebalance or reorder their automated scripts or automation, asks whether scripts conflict or overlap, which script should run first, or why a script never does anything.

Create new article/ad titles for a division by chatting, based on what is currently winning, and submit the approved ones either as title recommendations (SetNewArticleTitles) or, only when the user explicitly asks to launch them, directly for article and ad generation (SubmitTitlesForGeneration). Use when the user asks to create, brainstorm, expand or suggest article titles or title recommendations for a vertical, theme, or country, or to launch/generate ads for finished titles.

Scout and size a batch of new article/ad titles to test for a division from its recent performance data. It measures how new titles have been doing (hit rate, cost of a failed test, profit of a winner), works out how many extra titles to test so spend grows while ROI stays above a floor, decides where to place them (country x vertical x feed), writes titles that carry the DNA of what is winning and avoid what is losing, and submits them for review, for generation, or as recommendations. Use when the user asks to scout, plan or propose new titles or title tests, asks how many titles to test, wants to spread into new angles, or says something like "find me titles for OH in DE", "what should we test next", "prepare N titles for review".