Limitations & caveats

This page sets expectations. Skeleton Key is in beta, and the sections below cover what is honest to know before relying on its outputs.

Run timing

Generation is not instant. The Brand Profile takes roughly 2 to 3 minutes. A Research run dispatches five analysts in parallel and lands in the 3 to 4 minute range; the optional executive one-pager and narrative report add another 4 to 5 minutes if you choose to generate them. Strategy is the longest phase: the four strategists take 7 to 8 minutes, and the Strategy Summary adds another 2 to 3 minutes for 9 to 11 minutes total. The Media Plan adds another 6 to 8 minutes. End-to-end, going from a fresh brief through to a Media Plan typically completes in under 30 minutes (excluding optional Research reports). Turning an output into an interactive page or a slide deck is a separate step that takes a few minutes. These are live durations you will see on the platform today.

Document upload limits

Each file in the Documents library is capped at 10 MB. Five formats are accepted: PDF, DOCX, XLSX, CSV, and MD. Anything else (including .doc, .txt, .png, and .pptx) is blocked at upload validation before the file reaches storage. See Documents for full upload behavior and per-run document selection.

Known model behaviors

Outputs occasionally include odd phrasing, formatting drift, or factual hallucinations. Niche or obscure categories produce sparser outputs, and the agents will say so when evidence is thin rather than fabricate confidence. On a transient failure, the workflow retries once automatically after a 3-second delay. If the second attempt also fails, the run is marked failed and you can rerun it from the workflow detail page. Skeleton Key is optimized for English content, so non-English markets work, but nuance and cultural specificity may be weaker. Cited sources are attached to claims throughout the platform, and the right move is to validate them before acting.

Beta caveats

Brand Maps, the Wiki, and Chat are the newest surfaces and still maturing. Treat Brand Maps as a sense-making tool: the extraction model occasionally misses elements or surfaces edges that look weaker than they are, and lowering the confidence slider reveals connections the default threshold hides. Wiki pages refresh on a queue, so a page can briefly lag the newest outputs or documents behind it. Chat answers from those same pages, so verify anything load-bearing by opening the cited page and reading the evidence. Output hiccups happen too. A misplaced table cell or a tone drift mid-paragraph turns up from time to time, and refining the run usually resolves it. Server stability is good but not perfect during back-end maintenance windows, and a page reload typically clears transient UI issues.

What doesn't work today

• No scheduled or recurring runs. Every workflow is triggered manually.
• No multi-language UI. The product is English-only at this stage.
• No on-premise or self-hosted deployment. Skeleton Key runs as a cloud product only.
• No separate B2B and B2C agents. The same agents run regardless of campaign type; the brief steers them, and the one behavioral difference is audience sizing, which uses LinkedIn's data for B2B campaigns and Meta's for B2C. See Integrations.
• No performance data connections. LinkedIn is the one account you can connect today, and it is read-only audience sizing, not results ingestion. Skeleton Key does not read your analytics, CRM, or ad-account performance data; documents you upload (and the CSV you bring to a Marketing Mix Model) are the only first-party data it sees.