Skip to content

What you get

your-project/
├── docs/airgap/blueprint.json answers, tier, sizing, decisions
├── docs/airgap/transfer-manifest.md images, models, packages and settings to carry in
├── docs/airgap/blueprint.md the blueprint (under docs/superpowers/specs/ with superpowers)
└── AGENTS.md instructions for the agent that builds it inside the network

These files travel into the network together with the artifacts the manifest lists.

The state of the design: the answers, the computed tier and sizing, and one decision per layer with its rationale, the alternatives considered and an evidence label. See blueprint.json.

Written in the language of the interview, in ten sections:

  1. Context: organization shape and constraints.
  2. Tier, floors and overrides.
  3. Architecture overview with a component diagram.
  4. Layer decisions: choice, version, rationale, alternatives, evidence, caveats.
  5. Sizing: inputs, concurrency range, GPUs, servers, power, assumptions.
  6. Air-gap transfer plan.
  7. Data flow: request path and authentication path.
  8. Failure modes and operations, including offline upgrades.
  9. Verification plan: benchmarks to run and the list of estimate decisions.
  10. Out of scope and later layers.

The one list of what to carry in, written by manifest.py from the chosen components: container images for your hardware, model artifacts, packages, telemetry and other outbound calls to switch off before first start, files to stage that the software would otherwise download, per-tool skills worth carrying, and checksum and signature steps. Its header records a digest of the decisions, so lint notices when the manifest is older than the design. Carrying it in explains how it is used.

A delimited block that tells the agent inside the network to follow the blueprint, use the pinned versions, assume there is no internet, and propose any new dependency as a blueprint change instead of adding it. Running the skill again replaces the block.