The Instruction File Behind Our AI Content Agent

An AI content agent needs more than a prompt. Its instruction file defines the sources, boundaries, approval rules, and stop conditions that turn generated copy into a controlled content workflow.

Our content plan is set by people in Notion. On each drafting day, the agent identifies the assigned article, reads its brief and source draft, checks every claim against approved sources, prepares the English and Ukrainian versions, and sends both for review. It does not decide what deserves publication, and it does not treat its own draft as approved.

This article went through that workflow. Below is the working logic behind it, with sensitive operational details removed.

The instruction file is an operating contract

A prompt asks a model to produce an answer. An instruction file defines how recurring work should run.

For our content workflow, that distinction matters. The agent has access to a planned topic, a handover draft, product context, editorial rules, SEO data, public pages, and a publishing format. Those inputs do not carry equal authority.

The instruction layer tells the agent which source wins when two versions disagree. It also defines which actions require approval, which facts need current verification, and when the correct result is a blocker rather than polished copy.

That is what makes an AI content agent different from an occasional writing assistant. The agent carries a repeatable process forward using maintained context, approved tools, and a reviewable sequence of steps.

The public version of the file

The real working instructions are spread across a small set of operational documents rather than one oversized prompt. The public version below preserves their decision logic while removing private routing, credentials, connector details, and unpublished business information.

# Content agent operating instructions

## Assignment
Read the article assigned to today in the approved content plan.
Preserve its intended audience, topic, and business purpose.

## Source order
1. Current approved company and product sources
2. Current public website and published articles
3. The approved editorial brief
4. The task-specific handover draft
5. Search-planning data

If sources conflict, use the higher-authority source.
A keyword is never evidence for a product claim.

## Fact rules
Verify product details, prices, metrics, client references, dates,
quotes, and named tools before using them.

Do not strengthen uncertain evidence.
Do not fill missing facts with plausible language.

## Voice
Write for operators.
Use concrete language and short sentences.
Keep the human responsible for priorities and approval.
Describe autonomy as something each workflow earns.

## Delivery
Prepare one complete English article and one natural Ukrainian article.
Keep their publication metadata aligned.
Use English keyword planning only for the English version.

## Approval
Deliver drafts for human review.
Do not publish or schedule from a drafting request.

## Stop conditions
If a required fact, quote, metric, asset, or decision is missing,
state the blocker and identify the required owner or approver.

The useful part is not any single sentence. It is the order of decisions.

A human owns the plan

Our Notion content plan names the topic, audience, angle, publication date, and supporting material. The agent executes that plan instead of creating a new editorial strategy during each run.

This division prevents a common failure in automated content systems: confusing production capacity with editorial judgment. A system that can generate ten articles has not proved that any of those articles should exist.

People still decide what Octocrew needs to say, which audience matters, and which evidence is ready for public use. The agent handles the repeatable work around that decision: retrieving the brief, checking the required sources, structuring the article, preparing both languages, and packaging the result for review.

The same separation appears in our earlier account of a post drafted by the agent it describes. The agent produced the draft, but the brief, correction, and publishing decision remained visible human steps.

Sources have an explicit order

A source draft is useful evidence of intent. It is not permanent company truth.

That distinction protected this article from several outdated lines in its original handover draft. Product language, approval rules, and editorial conventions have changed since that version was written. The current approved context takes precedence.

The same rule applies to SEO. Our semantic core helps the agent identify that “AI content agent” is a relevant query. It does not authorize a new capability claim or force the phrase into every section.

The source order is simple:

1. Current approved product and company material defines the offer.
2. The live website defines currently published public copy.
3. The editorial brief defines blog structure and style.
4. The task brief and handover draft define the planned topic and examples.
5. Keyword data helps shape search intent and internal linking.

When two sources disagree, the agent does not blend them into a convenient middle. It follows the higher-authority source or stops for a decision.

Voice rules include rejected language

Positive guidance such as “be clear” or “sound credible” is too broad for dependable production. The agent also needs examples of language that should fail review.

Our content rules reject replacement framing, invented metrics, guaranteed outcomes, inflated autonomy claims, and generic AI hype. They require Octocrew to be described as a managed team of specialist AI marketing agents working inside the tools a company already uses.

The human remains responsible for priorities, judgment, and approval. New workflows start in approve-first mode. Reliable workflows may earn more autonomy within defined limits, while risky or hard-to-reverse actions remain under human control.

That operating model is explained in Autonomy Is Earned, Not Configured.

The file also carries editorial constraints that may appear small but protect consistency: US English for the English article, natural Ukrainian for the Ukrainian version, stable tag IDs, no invented bylines, and no hype added to make a weak claim sound stronger.

Bilingual publishing is one workflow

The English and Ukrainian articles are not two unrelated assignments.

They share a translation key, publication date, author, tags, CTA state, and draft state. The titles, descriptions, examples, and sentence structure can differ because each version must read naturally in its own language.

The English article uses the semantic core for search planning. The Ukrainian article does not inherit English keywords mechanically. It preserves the argument and evidence while using Ukrainian terminology and rhythm.

This pairing rule also prevents half-finished publication. Both versions stay in draft until the complete pair is ready and approved.

Stop conditions are part of quality control

An agent that always returns polished copy is unsafe.

Some assignments depend on an unpublished metric, a client quote, an asset, or a decision that the agent cannot verify. In those cases, a clear blocker is a better output than a finished-looking article.

Our drafting workflow instructs the agent to stop when:

• a required product fact has no approved source
• a public metric cannot be traced to its published context
• a client reference lacks clearance
• a quote cannot be verified word for word
• a planned image, screenshot, or example does not exist
• two authoritative sources conflict
• the requested action exceeds the approval attached to the task

These are not edge cases outside the writing process. They are part of the writing process.

The publishing gate remains separate

A completed draft is not a publication instruction.

For this workflow, the agent sends the paired drafts to the review channel. Alina can approve the copy, request changes, or hold the article. Publication remains a separate authorized action.

That separation gives reviewers a clean decision. They can assess the words without wondering whether a comment might accidentally trigger a public change.

It also matches the broader Octocrew model for managed AI marketing agents: define the workflow, restrict its permissions, inspect the output, and expand autonomy only when the evidence supports it.

Build your own instruction file

A useful first version can be short. It needs enough detail to settle the decisions that repeatedly slow down or damage your content process.

# {Brand} content agent

## 1. Assignment
What recurring work should the agent complete?

## 2. Audience
Who is the work for, and which reader takes priority?

## 3. Approved product description
Which paragraph may the agent rely on when describing the company?

## 4. Source hierarchy
Which sources are approved, and which source wins during a conflict?

## 5. Claim rules
Which prices, metrics, clients, quotes, and comparisons may appear?

## 6. Voice
Which examples match the brand?
Which examples were rejected, and why?

## 7. Workflow
Where does work begin?
Where is the draft delivered?
Who reviews it?

## 8. Approval
Which actions always need a person?
Which low-risk workflows may earn more autonomy?

## 9. Stop conditions
What missing information must produce a blocker?

## 10. Change ownership
Who may update these instructions?
How are approved corrections recorded?

Start with the decisions your team already repeats. Add a rule when it resolves a real ambiguity or prevents a real mistake. A long file filled with generic advice gives the agent more text, but not necessarily better control.

Turn your content process into an operating brief

If your team keeps re-explaining the same voice, sources, approvals, and client boundaries, book a discovery call. We will map the workflow and identify what the agent needs before it touches a live channel.

Frequently asked questions

What is an AI content agent?

An AI content agent carries a defined content workflow forward using maintained context, approved sources, tools, and review rules. It can retrieve a brief, prepare content, package the result, and route it for approval without requiring a new prompt at every step.

Is an instruction file the same as a prompt?

No. A prompt usually describes one task or requested output. An instruction file defines recurring operating rules, source authority, permissions, stop conditions, and approval ownership across many tasks.

Can a template make an AI content agent safe?

A template helps, but it is only one control. The workflow also needs narrow access, current sources, action records, human review, and explicit rules for missing or conflicting information.

Does the Octocrew content agent publish automatically?

Not in this drafting workflow. It prepares the paired articles and sends them for review. Publication requires separate authorization. Other reliable, lower-risk workflows may earn more autonomy within defined limits.

What should be redacted from a public instruction file?

Remove credentials, private routing details, unpublished metrics, confidential client information, and security-sensitive implementation details. Keep enough decision logic to make the example useful and honest.