Service Line

Talent Factory Build

Talent Factory Build Service Line

One-liner. Take a client's specific need ("we want an AI that does X for our team Y") and manufacture a digital talent through the 9-stage production line, ending with a deployable package the client runs themselves.

This is the factory's core product line. Each delivered talent is a foundry artifact — the client owns and runs it, the factory does not host. See feedback_delivery-model-foundry-not-hosted.

When to use this line

Use when the request is about producing a new digital talent (or a meaningful new capability of an existing one):

  • "Build an agent that extracts safety lockout sheets from PLB drawings" → custom talent
  • "We need an agent-ea variant tuned for Transgesco's data model" → talent variant
  • "Automate our weekly compliance report" → new talent
  • "Add a cadenassage L2 cross-plan capability to the existing extractor" → talent capability extension

Do not use this line for: documenting an architecture (→ Enterprise Architecture), auditing a process (→ Consulting), evaluating an off-the-shelf tool (→ R&D).

Inputs

The intake form (/request-create --service=talent) collects:

  1. Client + sponsor
  2. Use case (one paragraph)
  3. Target users + expected volume
  4. Inputs the talent will consume
  5. Outputs the talent must produce + format
  6. Existing artifacts to reference (similar talents, prior runs)
  7. Constraints (privacy, on-prem, local model required, etc.)
  8. Target deployment platform (Claude Code, OpenAI Codex, both, custom)
  9. AC + DoD draft (per feedback_document-before-shipping-standards)

Standard production process

The full 9-stage production line. Status: pipeline matures order by order — current STM order is at Stage 4 (Assembly).

1. Intake          →  2. Requirements    →  3. Design
4. Assembly        →  5. Test            →  6. QA Gate
7. Package         →  8. Deploy          →  9. Handover + Support

See /production-lines/digital-talent for the canonical stage definitions and gate criteria.

Deliverables (deployment package contract)

  • role.md — what the talent does, inputs, outputs, interactions
  • agent.md — Claude Code agent definition (or platform-specific equivalent)
  • processes/ — documented processes the talent follows
  • MODEL-CONFIG.md — model IDs + sweep script (per project_model-config-pattern)
  • README.md — client-facing run instructions
  • tests/ — regression tests
  • Maya bundle (when applicable) — RAG-as-wiki conversational KB scoped to this talent (project_maya-rag-wiki-pattern, WO-PROD-008)

Acceptance Criteria + DoD

Driven by the 9-stage gate criteria. Each stage has its own AC. Stage 6 (QA Gate) enforces the cross-cutting DoD.

Publishing target

Two-track:

1. Runtime — client platform. The talent runs on the client's own Claude Code / Codex installation. Factory delivers the package and supports first-run setup. Not hosted by the factory. LLM provider flexibility is part of the contract — the talent must run on at least Claude + OpenAI by default.

2. Documentation hub — JCT portail bundle. Each delivered talent gets a published documentation hub on the client's JCT subdomain alongside the runtime package: {client}.jacksoncreektech.ca/talents/{talent-name}/. The hub contains:

  • index.html — what the talent does, scope, runtime requirements
  • runbook.html — how to invoke, common workflows
  • architecture.html — interactive diagram of the talent (where it fits in the client's IT landscape, reuses the EA pipeline output)
  • changelog.html — version history
  • Maya conversational widget — embedded, scoped to this talent's corpus (per project_maya-rag-wiki-pattern)

Talents publish themselves. Same direction as Edward for EA: each delivered talent owns its own publish-to-JCT pipeline. The talent's deployment package includes a publish target config, and a publish.md skill the talent invokes after a release. Wiring spec: REQ-CONS-009 (Talent Factory bundle publication pattern).

Worked examples

Talent Client Status
agent-ea v1 STM Shipped (frozen pending v2 regen)
Cadenassage extraction (REQ-CONS-005) STM Pipeline B (DXF) live, 130 fiches générées
EA agent STM Pending handover (revenue blocker)
Maya (RAG-as-wiki bundle) All clients WO-PROD-008 in flight

Lead role

Production line operators — coordinated by Ivan (Infrastructure Engineer) for setup, Quincy (QA) for gates, Marcel (Methodology) for stage compliance. No single "talent factory manager" yet — the production line operates as a relay across the 9 stages.

Source-of-truth links

  • Production line: production-lines/digital-talent/
  • Methodology: departments/methodology/
  • QA: departments/quality-assurance/
  • Frameworks: see framework-library (TFD-010)

Status

3 talents shipped, WO-005 in Stage 4. Service line is active and proven, but production-line documentation is still being formalized (Marcel's ongoing rollout).