Content Creator

Content Creator helps people produce consistent, reviewable content without giving an AI control of their identity, opinions, or publication decisions.

Each author works in a small, independent repository containing their voices, agents, learning, drafts, and approved content. This Core repository supplies the reusable workflow, validation, and safety boundaries.

Here, voice means the author's written communication style: observable patterns in how they shape text. It is not a model of the whole person and does not claim to capture their identity, personality, beliefs, expertise, or inner character. Authorised transcripts may provide supporting evidence, but the resulting voice guides text creation.

Read Why not just use a chat app? for the design rationale.

Content Creator Core is distributed through PyPI as content-creator. Author workspaces pin an exact package version so installation and upgrades remain reproducible.

The quickest way to start

Give this request to Codex, Claude Code, or another coding assistant with terminal and filesystem access:

Use Content Creator Core to create a thin content workspace for me. Do not clone or copy Core. Follow its workspace-creation guide, ask me for the author and content choices you need, install the workspace, and validate it.

The assistant will ask about the author, intended content, writing evidence, and model provider. It will then create a separate workspace pinned to an immutable Core release.

Prefer doing it yourself? Follow Create a thin content workspace for the complete terminal instructions and configuration options.

How it works

The author first establishes an approved voice, then uses it in a repeatable content loop. Core's rules remain authoritative underneath every agent.

Voice setup

flowchart TD
    A["Create or open an author workspace"] --> B["Inspect persisted state<br/>with Start or Overview"]
    B --> C{"Choose a voice route"}
    C -->|"Authorised writing"| D["Measure a linguistic signature<br/>and build a voice candidate"]
    D --> E["Voice Analyst"]
    E --> F["Profile Critic"]
    F --> G["Voice Evaluator"]
    C -->|"No writing yet"| H["Use the neutral<br/>Clear Professional Starter"]
    G --> I{"Human review"}
    H --> I
    I -->|"Approve"| J["Activate an immutable voice version"]
    I -->|"Revise"| C

Content loop

flowchart TD
    A["Describe the content in ordinary language"] --> B["Coordinator resolves voice, pack,<br/>provider, perspective, and research route"]
    B --> C["Briefing Agent"]
    C --> D{"Research required?"}
    D -->|"Yes"| E["Researcher and research approval"]
    D -->|"No"| F["Writer"]
    E --> F
    F --> G["Critic and deterministic validators"]
    G --> H["Attribution Reviewer when required"]
    H --> I{"Human editorial review"}
    I -->|"Revise"| B
    I -->|"Approve"| J["Save inside the author repository"]
    J --> K["Learning Extractor updates<br/>only the active voice's learning"]
    K -. "future request" .-> B

Throughout both flows, persisted files—not chat history—hold workflow state. Core does not silently change a voice or perspective, invent identity or evidence, add research to a no-research route, or publish externally.

1. Create an author workspace

The generated repository belongs to the author. It keeps their editorial material separate from the reusable engine and can have its own writer, researcher, critic, policies, and learning.

Core is installed as a versioned dependency; its source code is not copied into each content repository.

2. Choose a voice route

The author chooses one of two starting points:

• Use previous writing to build a reviewable voice candidate from authorised documents or URLs.
• Start without previous writing using the neutral Clear Professional Starter, which does not claim to imitate the author.

No candidate becomes active without human approval. A new subject also does not automatically require a new voice; it may be better represented as a separately governed perspective.

See Voice onboarding for the commands and lifecycle, or How voice is derived for the underlying analysis and safeguards.

Statistical voice evidence

For a source-derived voice, Core calculates a deterministic linguistic signature from the authorised corpus. It records per-source measurements and descriptive statistics for rhythm, structure, stance, punctuation, and lexical diversity, separated by register where possible. These measurements are evidence for human review: they are not generation targets, personality inferences, forensic authorship attribution, or proof that a feature is unique to the author.

Read the statistical and linguistic voice framework, or inspect a built candidate with content-creator voice signature <voice-id>.

Core can also calculate an optional statistical_voice_score for a draft using either deterministic distribution comparison or an explicitly trained ML classifier. Scoring is disabled by default and remains critic-only advisory evidence: it is not an authorship probability, writing target, quality-gate input, or publication gate. Automatic workflow scoring additionally requires an eligible long-form content pack. The explicit voice score command remains available for deliberate, ad hoc assessment of any sufficiently long text.

3. Ask for content naturally

Open the author workspace in a supported coding assistant and describe what you need:

Write a short LinkedIn post explaining why calculus matters to sixth-form students. No external research is required.

The Content Creator Coordinator reads the workspace state, proposes the appropriate voice and format, follows the required research and review checkpoints, and preserves the run artifacts.

Automation and agent hosts can attach an --idempotency-key to a run. An equivalent retry returns the existing run_id and state instead of executing the content route twice; conflicting reuse fails safely. Intentional revisions use a new key plus --parent-run.

The result always comes back for human review. “Publish” means saving an approved copy inside the author repository; Core does not post to external platforms.

Recovered Core diagnostics remain out of the editorial conversation until the author approves the piece. At that publication boundary, Core presents one sanitised, consolidated support candidate and keeps external issue submission under explicit human control. Fatal Core diagnostics are surfaced immediately.

See Content Creator Coordinator for conversational and direct terminal use.

What the author controls

The author workspace owns:

• voice evidence, candidates, approvals, and immutable versions;
• perspectives and their provenance;
• repository-specific agents and editorial policies;
• repository-wide and voice-scoped learning;
• research, drafts, critiques, and run history; and
• approved content.

Core owns the reusable orchestration, schemas, provider adapters, validation, checkpoints, safety rules, prompt composition, and workspace generator.

Important boundaries

• The author remains the final editorial authority.
• Core does not invent personal experience, beliefs, facts, or voice evidence.
• Voices and learning are isolated from one another.
• No-research requests remain no-research.
• Voice activation and repository publication require explicit approval.
• External publication is not supported.

Guides

Use the task-oriented documentation index to create a workspace, manage a voice, create content, maintain a downstream repository, or develop Core. Detailed terminal procedures remain in the focused guides.

Work on Content Creator Core

Clone this repository only when you want to inspect or change the reusable engine. The Core development guide covers installation, architecture, testing, and releases. The engineering standards define the required quality, compatibility, security, and release controls.

External code contributions are not currently accepted. Bug reports and feature requests are welcome through GitHub Issues; read Contributing first. Report security-sensitive issues through the private process in the security policy.

Licence

Content Creator is free and open-source software licensed under the GNU Affero General Public License, version 3 or later (AGPL-3.0-or-later). Commercial use is permitted subject to the licence terms. See Licensing for a plain-language overview.

Copyright © 2026 Bharath Vadhoola