What this system is designed to control
Abstract
Amazon work becomes unreliable when facts, assumptions and performance data are mixed in the same document. This suite separates the job into six skills with explicit handoffs. Each stage consumes a defined evidence package, writes a defined artifact and preserves unknowns. The objective is not autonomous account control. It is a reviewable operating record that a brand team can inspect before copy is published, bulk files are uploaded or optimization decisions are applied.
1. Six skills, one evidence chain
The router starts with the earliest missing artifact instead of jumping directly to copy or advertising. A later skill may use an earlier artifact, but it should not silently rewrite the source evidence.
01 · amazon-listing-collector
Collect product evidence
Normalizes owned listing material, product facts, comparison inputs and explicit unknowns into a traceable source pack.
02 · amazon-review-analysis
Extract customer language
Turns review evidence into recurring needs, objections, use contexts and wording patterns while retaining source boundaries.
03 · amazon-keyword-library-builder
Build a governed keyword library
Maps product terms, use cases and customer phrases to intended placements instead of treating every discovered term as publishable.
04 · amazon-listing-generator
Generate a reviewable Listing draft
Produces structured copy from approved facts and keyword evidence, with claims and missing proof kept visible for human review.
05 · amazon-fad-initialization
Prepare paused ad drafts
Builds a launch structure, budget allocation and bulk-file draft. Every generated bulk row remains paused until reviewed.
06 · amazon-ad-report-analyzer
Turn one report into an action register
Analyzes one supported advertising report at a time and writes decision evidence, action candidates and a manifest—without writing back to the account.
2. The handoff contract matters more than one perfect prompt
Each stage should preserve three layers: observed evidence, an explicit interpretation and a proposed action. That structure lets a reviewer challenge the interpretation without losing the original source.
owned / permitted source material
-> evidence pack
-> review findings
-> governed keyword library
-> Listing draft + gate
-> paused advertising draft
-> one-report analysis + review queueDecision rule: if a required upstream artifact is missing, the system should stop at that boundary or label the assumption. It should not silently invent product facts, search-volume data or live performance.
3. Food-storage-container example: a draft, not a result claim
The included example demonstrates how artifacts move across the suite. Its numbers describe readiness and allocation inside a draft package. They do not describe sales, CTR, CVR, ACOS, ROAS or customer count.
All generated bulk rows remain paused. A reviewer must confirm product IDs, campaign structure, bids, budgets, negative candidates and marketplace requirements before any upload or activation.
4. Operational boundaries
- No account write access. The public method and packaged skills create files for review; they do not publish a Listing, upload a bulk file, change bids or activate campaigns.
- No invented commercial proof. A plan score or budget draft is not evidence of revenue, conversion or advertising efficiency.
- Negatives are candidates until approved. Search-term exclusions affect future reach, so evidence and reviewer ownership must stay attached.
- One advertising report per analysis run. This keeps column meaning, thresholds and output provenance inspectable.
5. Public method, protected implementation
This article is the public canonical explanation. The file tree, validation notes and downloadable suite are kept on a password-protected code page. The GitHub repository is private.
Mirrors: DEV Community · Medium. Both editions point back to this canonical page.
Inspect the implementation
Password required for source documentation and the packaged download.