How does BYO-LLM Guided Build actually work, step by step?

SaSame's normal customer flow performs no LLM calls and never modifies creator code — this is a machine-enforced invariant (mcp_network.mcp_builds.llm_calls = 0), not a marketing description. Concretely: at each build step, SaSame hands the creator a block of exact instructions; the creator pastes that block into their own LLM (Claude, ChatGPT, or whatever they already use under their own subscription); the LLM produces artifacts; the creator submits those artifacts back to SaSame; SaSame checks them with fixed rules — never a second LLM grading the first one's work.

A from-scratch build runs 7 fixed steps. Shorter entry points (existing repo, near-finished build) skip the earlier ones.

Each step is validated against a fixed JSON Schema (step_id pattern, a closed set of check kinds: min_length, max_length, regex_every_line, min_lines, contains_all, const_true, digit_min) — the recipe itself is versioned, so the exact rules a creator is held to at any moment are pinned and auditable, not ad hoc.

Real instructions text from step BLUEPRINT-01 — this is verbatim what a creator receives and pastes into their own LLM:

The creator's LLM returns two artifacts (product_definition, tool_inventory); SaSame checks them deterministically — length bounds, one tool name per line matching ^[a-z][a-z0-9_]*$ — and either advances to the next step or returns a plain-English failure_repair with what to fix.

Every single step in the schema carries no_secrets: const true. BUILD-04's deploy-readiness manifest is environment-variable names only, never values — every submitted string is scanned for secret-shaped values and rejected if one is found. This is enforced at the schema level, not left to reviewer judgment.

BLUEPRINT-02 must cover all of: purpose, target_user_or_agent, data_api_source_and_rights, minimal_tool_set, non_goals, tools, auth_and_effects, rate_limits, errors, hosting_choice, responsibility, distribution_metadata, observability_metadata. Every tool entry needs a name, description, input_schema, output_schema and example. The responsibility section explicitly assigns versioning, support and repair to the creator — SaSame's inspection and observation stations measure the result, they do not take over ownership of it.

BUILD-03's test run is explicitly self-reported: the creator runs their own suite locally and submits a SHA-256 of the output plus a pass count. These are the creator's own results — independent, external verification only happens later, at the Inspection station, which requires a real external tool call (not a local/self-test) before a build can ship. See Repair & reverification for how that later, stronger check works.

Concurrent guided-build slots are plan-bounded, not registration-bounded: Free Listing has zero (list/claim/observe only, no build capability), Factory gets 1 concurrent build, Factory Pro gets 3, Factory Team gets 10. See Pricing for the full capability table.