Deck analysis
The nine analysis panels: what each one is asked, what it can see, and why this surface deserves the most scepticism.
The analysis tab runs nine independent model calls over your deck and renders each result as a panel. It reads your whole list at once, which no other surface does, and that is what it is good at: telling you what the deck is trying to do and where the plan is thin.
It is also the surface with the least retrieval behind it, so its output is worth reading as an informed second opinion rather than as a checked result.
The nine runners
| Panel | Question it answers |
|---|---|
| Staple Coverage | Are there enough ramp, draw and removal effects |
| Top Performers | Which five cards punch above their weight here |
| Recommended Cuts | Which five cards are weakest or most replaceable |
| Cohesion | Do the cards work together as one strategy |
| Power Level | How strong is this, and does it match your stated target |
| Explosivity | How fast are the best draws |
| Design Alignment | Does the build deliver what the Design tab says it should |
| Ramp Analysis | Is the ramp the right quality and cost for this deck |
| Overall Summary | An executive summary across the other eight |
Each runs as its own Sonnet call with its own prompt and a JSON response shape. Overall Summary runs last and receives the other eight results.
What each runner can see
One deck summary, built once and reused as a cached system block so later runners pay roughly a tenth for those tokens. It contains:
- Commander and color identity.
- Total cards, average mana value, land count, tagged ramp and draw counts, and the on-curve cast rate.
- The full card list, with quantity, mana value, type line, tags and complete oracle text for every card.
That is everything. The runners have no tools. Unlike chat, they cannot search for a card, pull a rule, or look anything up mid-answer. Everything they say about your list is grounded in the summary above. Anything they say about cards outside it comes from the model itself.
What that means
Suggestions of cards you do not own are unverified. When a runner names a replacement for a cut candidate, that name comes from the model rather than from a search, so it has not been checked for existence, color identity or format legality. Treat those names as leads and confirm them in search or chat, both of which check.
Scores are judgements against a written rubric. The 1 to 100 power scale is defined in the prompt as described bands ("upgraded precon", "holds up in mid-power pods"), and the model places your deck among them. No other deck is used as a reference point. The band it lands in carries the meaning. The exact digit will move between runs on an unchanged deck.
The eight panels do not see each other. Each runs alone, so they can disagree. Only Overall Summary sees the rest, and it is the one that reads as coherent.
Interaction claims are not rules-checked. A description of how your combo works is written without consulting the rules text that chat can query. Take an interaction that matters to chat, where it can be looked up.
Caching
Results are stored with a hash of the deck list, so re-opening the tab shows the stored analysis rather than paying for a new run. Changing the deck changes the hash. The analysis you are reading may predate your most recent edit if the run has not been repeated.
Using it well
Take the structural reading, which is what a model over a full decklist is genuinely good at: whether the deck holds together, where it is thin, which cards are doing the work. Confirm the specifics, particularly card names it suggests you do not already own, somewhere that can check them.