How do you build content generation with AI agents?

Key facts

Endpoint	POST /v1/chat/completions per stage; JSON mode for briefs and outlines
Models	30+ models; route drafting, critique and tagging to different aliases
Structured briefs	JSON mode returns outline, audience, claims and required links as structured data
Fact checking	A verification pass compares draft claims against retrieved or provided sources
Pricing model	Flat monthly self-serve plans with unlimited fair-use usage
Free tier	Free plan with 2 free AI models, no card required
Governance	Scoped keys per pipeline and audit logs for generated artifacts
Roadmap	Images and batch endpoints are coming soon

TL;DR

• Split generation into stages; one prompt cannot draft, verify and edit reliably.
• Make the brief structured data so every stage validates against the same contract.
• Fact-check as a separate pass with sources, not as an instruction in the draft prompt.
• Route stages: cheap aliases for outlining and tagging, stronger ones for synthesis.
• Keep a human publish gate — the pipeline drafts, editors decide.

How it works, step by step

1. Define the brief schema: audience, goal, outline, required claims, sources and constraints.
2. Run a research pass to collect source material and record it alongside the brief.
3. Draft with a strong alias, requiring inline references to the provided sources.
4. Run a verification pass that extracts each claim and marks it supported, unsupported or needs a source.
5. Edit with a separate pass for voice, length and structure against a written rubric.
6. Validate links, banned phrases and required elements programmatically, then send to human review.

Original data

Try it yourself

Open the AI agent prompt generator →

Architecture of a content agent chain

Five stages, each with distinct acceptance criteria:

• Research: gathers source material and records it with identifiers.
• Brief: JSON mode produces a structured plan — outline, audience, claims, links.
• Draft: writes against the brief with inline source references.
• Verify: extracts claims and checks each against the provided sources.
• Edit: applies the style rubric without introducing new claims.

Stages fail differently, and that is the point. A weak draft is a style problem; an unsupported claim is a trust problem, and it needs its own pass.

Implementation details

Practical choices that improve output and reduce rework:

• Keep source material in the prompt for the draft and verify stages; never ask the model to recall facts from training.
• Make the verify stage adversarial: instruct it to find unsupported statements rather than confirm them.
• Cap the edit stage with an explicit rule that it may not add claims, only rephrase and reorder.
• Use cheaper aliases for outlining, tagging and metadata, and stronger ones for synthesis and verification.
• Validate outputs programmatically: link checks, banned phrase lists, word counts and required sections.
• Log every stage's input, output and model so a bad published piece is traceable to a stage.

Evaluation

Score the pipeline per stage rather than as a black box:

• Brief adherence: did the draft cover the required sections and claims?
• Factual support: share of claims marked supported by a provided source.
• Voice match: human or rubric scores against approved examples of your style.
• Edit distance: how much an editor changed before publish — the practical quality metric.
• Publish rate: share of drafts published without being rewritten from scratch.

Keep a set of past briefs with known-good outputs as a regression suite, and run it when prompts or aliases change.

Limitations

Generation does not remove editorial responsibility:

• Models can state unsupported claims confidently; verification only works against sources you actually provide.
• There is no built-in web research tool, so source gathering stays in your pipeline or a connected tool.
• Voice imitation needs examples; a style guide alone rarely captures a publication's tone.
• Independently producing similar drafts risks derivative writing; editorial originality review remains human work.
• Images and batch endpoints are coming soon, so visual assets and bulk runs need separate tooling today.

Honest comparison

Capability	Content agent chain on Plugsky	Single-prompt generation	Human-only writing
Fact verification	Dedicated stage against sources	None or boilerplate instruction	Manual research
Voice control	Style examples plus edit rubric	Prompt description	Native
Model choice	Per-stage aliases among 30+ models	One model	Not applicable
Traceability	Per-stage logs with source ids	One output	Editorial notes
Cost shape	Flat monthly self-serve, unlimited fair use	Per-token or per-seat	Salaries

Frequently asked questions

Why not use one prompt for the whole article?

Because drafting, verification and editing optimize for different things. One prompt tends to produce fluent text with unsupported claims and inconsistent structure.

How do I stop the model inventing facts?

Provide source material in the draft prompt, require inline references, and run a dedicated verification pass that marks unsupported claims before editing.

Which model should draft versus verify?

Use a stronger alias for synthesis and a separate pass for verification. Outlining, tagging and metadata can run on cheaper aliases from the 30+ model catalogue.

Does Plugsky search the web for research?

No. Research gathering happens in your pipeline or a connected tool; Plugsky generates and verifies against the material you provide.

How do I measure content quality?

Track brief adherence, share of supported claims, rubric-based voice scores, edit distance before publish and publish rate.

Can non-writers use this pipeline?

With a human review gate, yes. The pipeline removes blank-page work and enforces checks, while editors keep judgment over claims and tone.

Can I test the pipeline for free?

Yes. The free plan includes two free models with no card, and the 14-day full-access trial allows comparison of drafting aliases on your own briefs.