Skip to content

AI in Dezycro

Dezycro uses AI at four key surfaces. All four are grounded in your real product context via RAG over ingested documents.

1. PRD authoring

The AI chat (web UI) and the /dezycro:author --doc prd skill (Claude Code) let you create PRDs conversationally. Describe what you want; the AI produces a structured document. Refine via follow-up messages.

Output PRD structure:

  • Overview / user stories / acceptance criteria
  • Pillars — top-level capability groupings (## Pillar 1: <name>) that the TRD must subsequently address by name

PRDs are not gated — finalize whenever you're happy.

2. TRD authoring

The TRD is the technical translation of the PRD. Same authoring surfaces (chat + /dezycro:author --doc trd). There's no required schema or format — see the TRDs guide for what tends to work well.

On APPROVE, Companion runs a light sanity check (placeholder tokens like TBD / TODO / FIXME, very-short drafts) — not schema enforcement; just a "did you mean to do that?" guard.

3. Test case generation

In LogicStudio, the AI generates test cases (which become journeys in the workbook) from the TRD. Coverage targets:

  • Happy paths — standard user flows
  • Edge cases — boundary conditions
  • Negative testing — invalid inputs, auth failures
  • Integration points — interactions between features

Each generated journey is grounded in a linked intent — the specific TRD/PRD statement it tests. This traceability is what enables coverage reports.

4. Workbook sweeps (Claude Code)

The Companion V2 plugin runs an LLM-driven sweep of your recent Claude Code session to extract decisions, issues, assumptions, and task-status hints, and auto-log them to the workbook. Silent by default; triggers on wrap phrases ("save progress", "/dezycro:sync") or after /dezycro:verify.

How RAG works

All four surfaces use retrieval-augmented generation:

  1. Your query/draft becomes an embedding (Cohere embed-v3)
  2. A hybrid (dense + sparse) search runs against your project's Qdrant collection
  3. Top-30 candidates re-ranked by Cohere rerank-v3
  4. Top-K chunks injected into the LLM prompt as grounding context

The result: AI output cites your actual product terminology, reuses existing flow patterns, and doesn't invent details that contradict what you've documented.

Embedding calls run on shared infrastructure and do not consume your tenant's AI token allowance.

Model selection

By default, all LLM calls go through Anthropic Claude (Opus 4.x family for high-stakes authoring, Sonnet 4.x for chat and lighter generation). Enterprise tenants with BYOK can route to their own provider (Anthropic / OpenAI / Bedrock / Vertex / Azure OpenAI).

All tenant-facing calls route through TenantLlmService, which resolves the right model + credentials per-tenant transparently.

Conflict detection

When your PRDs (or PRD sections) contain contradictory requirements, the AI flags conflicts with references to the specific contradicting passages. Resolve from the conflict review UI.

Iterative refinement

AI generation isn't one-shot:

  • Add detail to PRD sections via follow-up chat
  • Request additional test coverage for specific areas in LogicStudio
  • Re-run TRD authoring with --revise-existing to refine an approved draft

Tip

The best results combine AI generation with human review and domain expertise. Treat AI output as a strong starting point, not a finished product. The TRD quality gate and Companion's intent-violation tracking are there specifically to catch hallucination drift.