HL ConsultingKNOWLEDGE HUB

← Journal

Cobra · Product AI agent

Product AI agent — from one-shot script to dashboard CDC

Claude · HL ·
agent odoo-dev strat doc en-cours Cobra

Context / the need

June 2026. We just created the 36 Davis Acoustics variants (Krypton + Ariane ranges) in Odoo — see the dedicated entry. It worked, but it took time, it was technical, and it can't be a non-technical operator doing this alone next time.

The problem: creating a range in Odoo so far required a hand-written Python XML-RPC script. For Davis: 11 templates, 36 variants, 72 supplier lines (stock price and drop price, 2 per variant). Every brand started from scratch; every color-naming or SKU-format mistake was fixed after the fact.

What triggered the formalization: Davis = the first new brand created since the Cobra Odoo instance went to production. Big enough to make "let's encode this once and for all" a real question. It's the concrete realization of the automatic product-creation agent sketched in May.

What was done

The agent

A conversational AI agent, hosted in the internal Cobra dashboard. It accepts any source file (a photo of a price list, an Excel catalog, a PDF), analyzes it via the Claude API (Cobra key), asks the user the missing questions, shows a recap, waits for a confirmation, then performs the creations and updates in Odoo via XML-RPC.

4 supported actions:

  • Create a range (templates + variants + supplierinfo STOCK/DROP);
  • Update sale prices;
  • Update purchase prices;
  • Update sale + purchase at once.

What it never does without a human: activate products on Shopify. actived_in_shopify=False on every creation, no exception.

The dashboard

No new tool — we build on the internal Cobra dashboard, already in production, already accessible to the whole team. The agent page shows an upload zone (drag-and-drop), an action selector (4 buttons), a conversation area (multi-turn Q&A with Claude) and a downloadable final report. Why the dashboard rather than an Odoo wizard: zero training needed, a conversational interface usable by non-technical people, no dependency on the Odoo.sh release cycle.

The spec (CDC)

A spec (CDC-AgentProduitIA-…-juin2026.html) handed to the dev provider. It specifies everything a developer needs to implement the agent without coming back with questions: architecture (Flask/Node + Claude API + XML-RPC), Claude's JSON output format (status: questions | ready | error), a full system prompt with SKU/EAN/color/category/fixed-field conventions, error handling (4 cases), security. The "Cobra dashboard" HTML format is deliberately the same as the other specs — it's the intended template for the next brands.

Decisions and discarded alternatives

Design raised about ten structuring questions. The main ones:

  • Pair vs unit. Floorstanders and bookshelves sell per pair — the Odoo list_price is the pair price, the template name contains (la paire); centers are per unit. Observed on the Dynaudio Emit records before writing a single line.
  • SKU format. BRAND-MODEL-COLOR uppercase, short slug (e.g. DAVI-KRY10-NR). Mirrored on what exists; short = fewer typos.
  • Double purchase price. Two supplierinfo lines per variant, same supplier: SKU-STOCK (cheaper) and SKU-DROP (drop, pricier). The suffix makes the info readable directly in Odoo.
  • Color naming. Capitalize each word, / separator for two-tones (Vert / Chêne, Noyer / Blanc). Rule formalized after a first creation in the wrong case.
  • Language. French only; English commercial names translated at ingestion (Nordik → Blanc, Classik → Noyer, Technik → Noir).

Discarded alternatives:

  • Odoo wizard: too coupled to the instance, not accessible without admin rights.
  • Native Odoo CSV import: doesn't handle multi-attribute variants or the double supplierinfo line.
  • Case-by-case manual mapping: doesn't scale, no guaranteed consistency.
  • Odoo → dashboard SSO: different domains, impossible without dedicated infra → email check against Odoo res.users via XML-RPC.

Snags and how we fixed them

  • The brand field. x_studio_marque returned an error. Resolved with a fields_get() on an existing product → discovering product_brand_id.
  • The uniqueness check in the wrong order. The SKU/EAN check ran after the write at step 3 — the script found its own fresh records and blocked. Worked around by creating the supplierinfo lines in a second script.
  • Mis-named colors. First creation in vert-chêne, bleu-noyer… instead of the Cobra convention. Fixed via write() XML-RPC on the product.template.attribute.value; lesson encoded as a naming rule in the spec's system prompt.
  • The supplierinfo product_code. Three iterations (no suffix → SKU only → -STOCK/-DROP), 72 lines updated by script each time. Final version frozen in the spec.

Result

As of June 10, 2026:

  • Davis Python script: in production — 36 variants created, 72 supplierinfo lines, 0 duplicate EAN.
  • Product AI agent spec: written, handed to the dev provider.
  • Dashboard ↔ Odoo auth spec: written, handed over. Auth will mirror the list of internal users active in Odoo.
  • Dashboard (development): pending on the provider's side, ~3 days estimate.
  • One-shot script idempotency: known bug (the uniqueness check), non-blocking for the Davis use.

Next

  • The dev provider builds the backend + the conversation UI.
  • First test on a new brand (validating the system prompt in real conditions).
  • Version the spec in the Cobra repo (currently local only).
  • Assess whether the spec is enough as a template for the next brands, or whether the system prompt needs a per-brand calibration phase.