Content Factory

Русская версия

A Markdown-first, human-gated multi-agent workflow for producing long-form editorial content from source material or research.

Content Factory separates research, planning, writing, and whole-document editing into explicit roles. Each machine-produced stage leaves an inspectable file, human approval is explicit in the conversation, and writing cannot begin until that approval is given.

Project scope: this repository contains orchestration specifications, agent instructions, editorial policies, and templates. It is not a hosted application, an autonomous agent runtime, or evidence of unattended production volume.

Workflow

source file or topic
        |
        v
researcher (optional)  ->  research.md
        |
        v
planner                ->  plan.md
        |
        v
orchestrator review
        |
        v
HUMAN APPROVAL GATE
        |
        v
writer, block by block ->  draft.md
        |
        v
orchestrator/editor    ->  final.md

The orchestrator is the main agent context, not another isolated reviewer. It checks the plan, enforces the approval gate, reads the complete draft against the source and examples, and performs the final edit.

See the architecture notes for the artifact contract, stop conditions, and design rationale.

Execution model: sequential by design

The included write-* workflows are sequential. Research must finish before planning; the plan must pass review and human approval before writing; writer calls append blocks in order; final editing starts only after the draft is complete.

Multiple specialist agents do not imply parallel execution here. The repository does not include a parallel scheduler, job queue, retry engine, or observability layer. Independent research branches could be parallelized by a compatible runner, but that behavior is not implemented in this public version.

This constraint is intentional: downstream stages depend on explicit file artifacts, and later writer calls need the current draft to preserve continuity and avoid duplication.

What is included

Four reusable workflows demonstrate different input contracts:

• write-article: a product or service roundup based on a source list plus current research;
• write-essay: an argument-led piece based on user-provided material;
• write-overview: a topic-led explainer that starts with research;
• write-guide: a practical guide derived from a transcript or another spoken source.

The repository also includes:

• specialist role definitions in agents/;
• general writing rules and anti-patterns in skills/good-writing/;
• format-specific editorial policies in skills/redpolicy-*;
• project, content-type, policy, example-pack, and quality templates in templates/;
• empty input/ and output/ workspaces protected by .gitignore.

Quick start

1. Copy templates/project-profile.md and describe the audience, purpose, voice, fact policy, and prohibited patterns.
2. Choose one content type and fill in templates/content-type-profile.md.
3. Add two to five examples that the system should imitate under examples//.
4. Put the source material in input/, or choose a topic for a research-led workflow.
5. Give the relevant skills/write-*/SKILL.md specification to an agent runner that supports a main context and specialist agents.
6. Inspect research.md and plan.md, approve or revise the plan, then let the workflow continue to draft.md and final.md.

Runner-specific directory conventions are described in the Russian-language docs/install.md. A fuller Russian setup walkthrough is available in QUICKSTART.md.

How to evaluate a run

Model output is nondeterministic, so exact-text snapshots are not a useful test. The repository provides an artifact-based review protocol instead:

• confirm that every required artifact exists;
• trace current or risky claims back to research.md or the supplied source material;
• verify that no draft existed before explicit plan approval;
• compare every draft block with the approved plan;
• score structure, clarity, voice, and disclosure of missing evidence.

Use the repeatable rubric in docs/evaluation.md. The shorter operational checklist used by the Russian-language workflows is in templates/quality-checklist.md.

Design decisions

• Files are the interface. Agents pass durable artifacts instead of relying on hidden conversational memory.
• Planning is a hard gate. Structural mistakes are corrected before prose makes them expensive.
• Fresh facts are isolated. Research is written to disk with sources before planning or writing.
• Writing is incremental. Each writer receives the full plan and the existing draft, then appends only its assigned blocks.
• Final review stays in the main context. The orchestrator can compare the complete task history, evidence, examples, plan, and draft instead of applying a checklist to an isolated fragment.
• Project voice is configuration. Brand, audience, calls to action, and examples belong in user-owned profiles, not in the reusable core.

Evidence boundary

The public repository intentionally excludes private inputs, client or community material, production examples, and generated output. It demonstrates the reusable workflow and its quality controls. It does not claim business outcomes, a specific number of published pieces, autonomous operation, or parallel execution.

The reusable agent and skill specifications are currently maintained in Russian; this README, the architecture contract, and the evaluation protocol provide the English entry point.

License and author

MIT © 2026 Mikhail.

Author’s Telegram channel