Capsules
Talk through how the work is done. Save the background, working rules and sources in an editable capsule.
Keep the sources, qualifications and decisions behind your work connected. Package them as reusable context for your next AI conversation.
Reusable team context
Save what your team knows and how it approaches the work. Share the source editions, instructions and templates together so a teammate can use the same context in their own conversation.
Talk through how the work is done. Save the background, working rules and sources in an editable capsule.
Attach a template, a good example or earlier work to guide the format and level of detail.
Save the materials you use together. Share the exact versions with your team or export them as a JSON bundle.

A small knowledge graph links the records that matter to one question. A claim can point to the passage that supports it, the edition it came from, and the condition that qualifies it. A reviewer can follow those links and record where the evidence is incomplete before the team reuses the graph.
An ontology gives the team a shared vocabulary for record types and relationships. A semantic layer applies that vocabulary to a task. A source-checking layer follows claim-to-passage links. An update layer follows version changes and keeps the qualifications relevant to the current question.
A capsule can supply exact passages, the limits of their support and instructions to flag gaps. ALCE found that strong models frequently left claims without complete citation support. TACITUS uses claim-level review to test whether the cited passage supports the answer and whether important claims have citations.
Selecting the needed passages can reduce the number of input tokens. Liu and colleagues found that several long-context models performed worse when the relevant evidence was buried in the middle. Measure preparation, retrieval, generation and reviewer time together to find out whether the workflow saves time or money.
The reproducible assembly checks measure which source excerpts and qualifications survive context preparation, and how many tokens the result uses. Our earlier 48-response model comparison was mixed. Consistent gains in answer quality, cost and review time still need to be demonstrated.
Registration is open. AI generation is currently paused. Sign in with Google to use the App Library, save conversations, and attach or export exact versions of context.
Additional project
Compare how modeled audiences respond to policies, messages and product ideas. Inspect the assumptions and decide what to test with people.