# Compound Marketing > Compound Marketing - Skill: `everyinc/compound-marketing` (Agent Skill, multi-file: 13 files) - Install (CLI): `npx skillmds@latest add everyinc/compound-marketing` - Raw SKILL.md: https://api.skillmd.com/api/skills/everyinc/compound-marketing/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: everyinc (https://skillmd.com/u/everyinc) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/everyinc/compound-marketing --- # Compound Marketing Turn scattered company context into governed strategy and market action, then make the accepted work improve the next comparable run. Use one decision record across the work so human judgment survives every handoff. ## When to invoke - When Douglas asks to run Compound Marketing or use Marketing OS end to end - When a strategy, recurring marketing program, or GTM project has scattered sources and several decisions that must stay aligned - When substantial human edits or measured results should improve the next project - When the team needs evidence that a repeated workflow became faster or better Do not invoke for a single line edit, one narrow research question, visual production, or a task with no meaningful decision to preserve. ## Load the system Read: 1. `../../foundation/marketing-os/SKILL.md` 2. `../../foundation/marketing-os/references/golden-path.md` 3. `../../foundation/marketing-os/references/collaboration.md` 4. `../../foundation/marketing-os/references/knowledge-boundaries.md` Load `../../foundation/marketing-os/references/learning-loop.md` when an edit, ruling, result, or repeated pattern may change future work. ## Open the run Use `/.compound-marketing//` as the shared run directory. Before creating one, search `.compound-marketing/*/run.json` for a record with `run_status: open` and the same project and workflow. Resume it when found; do not create a duplicate run. Records marked `complete` or `abandoned` are terminal and must not be resumed. Create `run.md`, `run.json`, and `decision-record.md` there from the references in this skill. Do not store active project details in the Marketing OS repository. Pass the three absolute paths and current decision-record version in every specialist handoff. Establish: - The objective and intended audience behavior. - The project type: strategy, recurring program, or GTM. - The source owners for facts, decisions, protected language, structure, and approval. - The closest comparable workflow and available baseline. - The primary decision metric, review-burden measure, guardrails, and decision date. - The requested artifact and explicit exclusions. If the user has already supplied these answers, continue. Search supplied company sources before asking for anything discoverable. Ask one to three questions only when a missing human judgment could change the direction, approval chain, or definition of success. ## Carry one decision record Build `references/decision-record.md` before writing the deliverable. Give each locked decision a stable ID. Separate facts, decisions, protected language, evidence, inferences, open questions, and rejected directions. The decision record is the contract between stages. Every downstream artifact names the version it inherited and the decision IDs it preserved. A source conflict remains open until the named owner decides it. When the requested artifact is a 10- to 15-minute approval deck for a launch content ecosystem, route through `marketing/gtm` and its content-system approval-deck reference. The deck approves the system, production mandate, and remaining inputs; it does not reconstruct the strategy or ask leadership to approve every post. When an approved hero-film concept becomes a production treatment, route through `craft/compound-copywriting` and its video-treatment reference. Preserve the decision record's strategic tension, product mechanism, proof mandate, protected language, CTA, and constraints. Let the treatment solve character, scenes, visual action, humor, and timing without reopening the strategy. Record meaningful human changes by those operations, not by surface word count. ## Run the path Follow `../../foundation/marketing-os/references/golden-path.md`. It owns the stage requirements and specialist routing. At every handoff, update `run.json` with the artifact location, status, inherited decision IDs, protected-language IDs, open questions, and human ruling. The specialist produces the work; this skill keeps the governing record intact. ## Prove the system Read `references/proof-standard.md` before making an improvement claim. Resolve `` from this loaded skill's repository, then run: ```bash python3 /strategy/compound-marketing/scripts/check_run.py validate python3 /strategy/compound-marketing/scripts/check_run.py compare ``` Structural checks prove only that the run record is internally consistent. A named human must verify artifact fidelity before an operational or compounding claim. ## Use optional pods sparingly Read `references/optional-pods.md` only when three or more independent jobs can run from the same approved decision record. Keep the user in plain English. Agent count is never a success metric. When that reference's gate clears, load `../marketing-pod/SKILL.md` for the authoritative pod contracts and scorecard. ## Dependencies - `foundation/marketing-os` - `strategy/program-brief` (routed) - `strategy/one-pager` (routed) - `marketing/gtm` (routed) - `craft/compound-copywriting` (routed for related copy sequences) - `craft/editing` - `marketing-science/measurement` ## Quick checklist - [ ] The run has an objective, baseline status, metric, and decision date - [ ] Facts, decisions, protected language, and open questions are separate - [ ] The human approved the governing decisions - [ ] Downstream work names and preserves inherited decision IDs - [ ] Human edits and results were compared with the original inputs - [ ] Learning was classified before any durable file changed - [ ] Improvement claims use comparable prospective runs and pass the proof standard - [ ] Design, asset production, and channel automation remained outside V1