AI-generated text sounds like AI. It defaults to a particular cadence, vocabulary, and level of formality that has nothing to do with your company. Without explicit guidance, every output reads the same — regardless of whether it’s a marketing email, a legal memo, or a support reply.
JieGou’s brand voice and governance controls solve this at the platform level. Define your voice once, and every recipe execution automatically applies it.
Brand voice profile
A brand voice profile has five configurable dimensions plus reference material:
| Dimension | Options |
|---|---|
| Tone | formal, casual, technical, friendly, authoritative, conversational |
| Sentence length | short, medium, long |
| Jargon policy | avoid, moderate, embrace |
| Audience level | general, professional, expert |
| Custom instructions | free-text, up to 2,000 characters |
You can also attach up to 5 reference writing samples (1,000 characters each) that demonstrate your preferred style. The model uses these as few-shot examples when generating output.
Department-level overrides make this practical for real organizations. Marketing can be casual and conversational while legal stays formal and authoritative. Each department can override any dimension of the account-level default. The resolution hierarchy is straightforward: department override takes precedence over account-level default, which takes precedence over none.
How injection works
Brand voice is not a suggestion. It is automatically injected into every recipe execution as a system prompt section, wrapped in <brand_voice> XML tags with all configured dimensions included.
The injection happens inside the recipe executor. Workflows inherit it automatically — every recipe step in a workflow gets the same brand voice treatment. No per-recipe configuration needed.
The resolution order at execution time:
- Check for a department-level override matching the executing user’s department
- Fall back to the account-level default
- If neither exists, execute without brand voice
Fail-safe behavior: if Firestore is unreachable when the executor tries to load the brand voice profile, execution continues without it. A missing brand voice is better than a failed recipe.
Glossary
The glossary enforces consistent terminology. Each entry has five fields:
| Field | Purpose |
|---|---|
| Term | The word or phrase to match |
| Definition | What it means in your organization |
| Preferred usage | How it should appear in output |
| Context | When to use it (and when not to) |
| Department | Account-wide or department-specific |
When a user submits a task, word-boundary matching scans the input for glossary terms. Only matching terms are injected into the prompt — this keeps the context window lean instead of dumping every term into every execution.
Matching terms are formatted as a <glossary> XML block with a 500-token budget. If matching terms exceed the budget, they are prioritized by relevance to the input. Department-scoped terms for the executing user’s department are included alongside account-wide terms.
Prompt fragments
Prompt fragments are reusable prompt components that you reference in recipe templates with {{fragment:name}} syntax. Think of them as shared building blocks — a compliance disclaimer, a formatting checklist, a standard intro paragraph — that multiple recipes can reference without duplicating text.
Each fragment has:
- Name — the reference key used in
{{fragment:name}} - Category — intro, checklist, format, disclaimer, instructions, rules, context, or other
- Content — up to 10,000 characters
- Department scope — account-wide or department-specific, with optional filtering
- Usage count — auto-incremented each time the fragment is resolved during execution
The UI includes a copy-to-clipboard button so you can paste the {{fragment:name}} syntax directly into recipe templates.
At execution time, fragment references are resolved by replacing each {{fragment:name}} token with the fragment’s content. This happens after brand voice injection but before variable substitution and context assembly.
Context assembly order
Understanding what the model sees matters for debugging. Here is the full assembly order at execution time:
- Industry system prompt — the base prompt for your account’s industry
- Brand voice system prompt — tone, style, and reference samples
- Fragment resolution —
{{fragment:name}}tokens replaced with content - Variable substitution — user-defined variables filled in
- Context sections — glossary terms, few-shot examples, RAG documents
- User task prompt — the actual input from the user
Each layer builds on the previous one. If something in the output feels off, you can inspect each layer independently to find where the problem is.
Governance features
Audit logging
JieGou tracks 120+ distinct action types across the platform — including brand voice profile changes, glossary edits, fragment modifications, recipe executions, and workflow modifications.
Each audit event stores:
- Who — the user who performed the action
- What — the action type and affected resource
- Before/after snapshots — full JSON of the resource state before and after the change
- Timestamp — when the action occurred
Three logging modes handle different reliability requirements:
| Mode | Behavior | Use case |
|---|---|---|
| Fire-and-forget | Async write, no execution delay | Default for most actions |
| Synchronous | Blocks until the audit record is persisted | Compliance-sensitive operations |
| Compliance-aware | Auto-dispatches to the appropriate mode based on compliance level | Set-and-forget configuration |
Retention policies are configurable per account, from 30 to 365 days. Audit data can be exported to CSV or JSON with filters on action type, user, date range, and resource.
Compliance levels
Three presets match common organizational requirements:
| Level | Retention | Logging | Additional |
|---|---|---|---|
| Standard | 30 days | Async | Default for all accounts |
| Enhanced | 90 days | Async | Forced reasoning traces on every execution |
| Strict | 365 days | Synchronous | Immutable run history, required approval gates |
Strict mode means every audit write blocks execution until persisted, run history cannot be deleted, and workflows with approval steps require sign-off before proceeding. This is designed for regulated industries where auditability is a legal requirement.
RBAC for brand voice
Access to brand voice controls follows JieGou’s role-based access model:
| Action | Required role |
|---|---|
| Edit brand voice profile | Owner or Admin |
| Edit glossary terms | Editor or above |
| Edit prompt fragments | Editor or above |
| Read brand voice, glossary, fragments | Any member |
This means your content team can manage glossary terms and prompt fragments without needing admin access, while brand voice profile changes — which affect every execution across the account — require elevated permissions.
Availability
Brand voice profiles and glossary are available on Pro plans and above. Governance features — audit logging, compliance levels, and retention policies — are available on Team and Enterprise plans. Prompt fragments are available on all paid plans.