Start with the published instruction file
Give your project-aware agent the 1view authoring instructions. The single file includes the analysis process, worked investigations, file contract and offline tools. The authoring page provides a prompt and an optional downloadable kit.
Specify the source revision, directories, allowed access and output location. The instructions prohibit deleting, moving or overwriting the analyzed project’s files. Intermediate work belongs in the designated run folder. Analysis access is not permission to deploy services, change data or expose secrets.
Build an inventory before a story
The agent should account for source files, entrypoints, components, hosts, interfaces, storage and external systems before deciding what the product looks like. It should reconcile migrations and datastore definitions rather than taking a single convenient schema as complete.
Unsupported source families, inaccessible infrastructure and restricted files should become explicit gaps. This keeps an attractive diagram from hiding the boundaries of the investigation. The public reader cannot independently inspect the sources to repair those omissions.
Deepen the analysis in revisitable stages
The instructions organize work into eight stages and bounded, resumable tasks. Early structural findings are refined as the agent discovers connections, behavior, requirements and evidence. Later findings can reopen earlier conclusions and invalidate dependent completion checks.
For each component, investigate its responsibility and execution context. For each connection, inspect endpoints, contract and failure behavior. For each trigger, follow success and relevant adverse paths. The resulting views should explain the discovered system, rather than driving the agent to invent convenient connections.
Keep evidence and uncertainty together
Require source references and a recorded revision for material claims. Distinguish source-inspected facts, contracts, proposals, unverified findings and synthetic examples. A requirement describes what should hold; its existence is not proof that the implementation satisfies it.
Ask the agent to record missing deployment visibility, unresolved ownership and unsupported analysis. When time or access is insufficient, it should produce an honest partial result. Smaller models can benefit from bounded tasks and explicit guidance, but the process does not guarantee exhaustive semantic analysis.
Validate, inspect and hand off one file
The agent should run the supplied structural and analysis checks before delivering one .ospec. These checks can catch broken references, missing accounting and inconsistent declarations. They cannot prove that every interpretation is correct.
Open the file in 1view and sample real questions across architecture, data, flows and features. Compare several important answers with their cited sources. Preserve the file and unresolved questions with the source revision so another person or agent can resume the investigation without starting from a blank explanation.
Explore a fictional example, or open your own .ospec file locally.
Open 1view →