SEO helper: an agent that drafts your blog posts

Last updated: 2026-09-06

By the end of this runbook you will have an agent that researches a keyword, writes a blog post in your voice, draws its own figures, opens a draft pull request on your repository, and emails you the link. It runs on a schedule you choose, and it never merges anything. It takes about twenty minutes, most of which is writing the description of your own site.

This template packages the SEO Agent Pack by Nicolas Lecocq, MIT licensed. The method is his. This runbook covers running it on Orca.

Before you start

A GitHub repository

The repo your blog lives in, on an account you can authorize. Required. Without it the agent has nowhere to deliver and will stop rather than write something it cannot hand over.

A DataForSEO account

Where the real keyword data comes from: search volume, difficulty and cost per click. Pay as you go, and a small deposit lasts months. Set it up before you start, because the agent will not run without it.

You do not need to install anything, and nothing runs on your own machine.

Step 1: Open the template and copy it

Open the share link

Open the SEO helper template.You do not need an account to look. The page shows the full description and, beside it under the Copy button, a What gets copied panel listing everything the template installs: one scheduled run, two agents, one pod, and one skill.

Press Copy template

If you are not signed in yet, Orca sends you to sign up or sign in first, and creates your workspace if this is your first visit. When that is done it brings you straight back to this same template page rather than dropping you on a dashboard.Press Copy template again once you are back. The first press only started the detour.

Review the copy plan

A plan appears listing every asset, its kind, and the name it will take in your workspace.On a new workspace nothing clashes and you can install as is. If a name is already in use, that row is flagged name taken and Orca proposes a free one ending in -copy. You can type your own name instead, or untick the row to skip it. Nothing you already own is ever overwritten.

Install

Press Copy selected assets.Open Agents and you will see seo-writer and seo-media. Open Skills and you will see seo-engine. They are yours now, and every one of them is editable.

Step 2: Connect GitHub

The agent needs two connections before it will run: GitHub, so it has somewhere to deliver, and DataForSEO, so it picks keywords from real data. Start with GitHub.

Open the writer agent

Go to Agents and click seo-writer. It opens as a panel showing its capabilities, its skill, and its connected apps.Fresh from a copy it reads Needs setup, with a banner headed Finish setting up this agent. The banner is honest about the consequence: runs are refused until the required items are resolved, and nothing is charged for a refused run.

Press Connect on the GitHub row

The banner names each unresolved dependency and puts a Connect button beside it. That takes you to Connected apps with the right app already in focus.

Authorize the account that owns your blog repo

A GitHub consent screen opens. Authorize the account or organization that owns the repository you want posts delivered into.

Check it went active

Back in Orca the connection should read ACTIVE. If it reads INITIATED the popup did not finish; if it reads FAILED or EXPIRED, disconnect and try again.The agent’s Dependencies tab is the place to confirm. It lists everything the agent leans on outside its own prompt, healthy rows included, with a status against each.

The agent never goes looking for a repository. It only writes to the exact owner/name you give it in Step 4. Connecting GitHub grants access; naming the repo is what points the run at a target.

Step 3: Connect DataForSEO

This is where the keyword numbers come from. The agent is not allowed to invent search volumes, so without this connection it has nothing to choose a topic with, and the run is refused before it starts.

Create a DataForSEO account

Sign up at dataforseo.com and top up a small amount. It is pay as you go and keyword lookups cost a fraction of a cent, so a few pounds goes a long way.

Connect it in Orca

Back on seo-writer, press Connect on the DataForSEO row of the setup banner, and enter the credentials from your DataForSEO dashboard.

Confirm the agent is ready

With both connections active, the agent’s status changes from Needs setup to Ready and the setup banner disappears.

Step 4: Write your site prompt

This is the step that decides whether the drafts are any good, so it is worth ten unhurried minutes. The agent starts every run knowing nothing about you. That is deliberate: it is what lets the same pair of agents write for several different sites without the facts bleeding together. Everything it needs comes from a prompt you write once and attach to the run.

Open Prompts

Go to Prompts and create a new one. Name it after your site, for example example-com-site-prompt.

Fill in the three facts that only you know

Your domain, your repository as owner/name, and three to five subjects you want to be known for. Everything else has a sensible default or can be discovered inside your repo.If your blog is plain Markdown files in one folder, this is genuinely short:

Write and deliver one SEO post, following your seo-engine skill and its
run protocol. Pick the keyword yourself from the lanes below unless I
name a topic.

## Domain and repo
Domain: example.com. Repo: acme/website, base branch main.
Posts live at src/content/blog/, format md.
Sitemap: https://example.com/sitemap.xml

## What we sell, and who reads this
We sell invoicing software to freelance designers.
Readers are self-employed designers who bill by project.

## Allowed lanes
1. Getting paid on time: chasing, deposits, late fees
2. Pricing creative work
3. Bookkeeping for one-person studios

## Forbidden lanes
Generic business advice, crypto, AI news.

## Voice
Practical and direct, first person plural, no hype. Contractions always.

## Delivery
Branch seo/<slug>, PR titled [SEO draft] <title>, label seo-draft, base main.
Draft PRs only. A human merges. Never push to main.

What ours actually looks like

Ours is attached to the scheduled run as a prompt document. It’s longer, because our blog is a Vite React app where every post is a TSX component. The prompt carries our repo layout, design system and style bans:

If the runtime presents it under a `Context` or configuration header, treat
the site facts below as authoritative operator-supplied input for this run;
do not ask me to repeat them in the run message.

Write and deliver one SEO post, following your seo-engine skill and its run
protocol. Pick the keyword yourself from the lanes below unless I name a topic.

## Domain and repo

Domain: orcapods.ai. Repo: okikorg/orca, base branch main.

Posts are **not Markdown or MDX**. The blog is a Vite React multi-page app and
each post is a TSX component. Everything lives under `landing/`. Adding a post is
three steps plus assets:

1. A `BlogPostMeta` record in `landing/src/blog/posts.ts`. SEO posts must set
   `featured: false`, `category: 'Technical'`, `heroVersion: 1`, and both
   `heroBarLabel` and `heroBarTag` written for the post.
2. The article component at `landing/src/blog/<slug>.tsx`, built with
   `BlogArticleTemplate`, including TOC, FAQ JSON-LD and `sourceNote`. Use
   `BlogFigureLive` and `BlogCodeBlock` for figures.
3. One registry line in `landing/src/blog/registry.ts`.

Then put images in `landing/public/blog/` and append the canonical URL to
`landing/public/sitemap.xml`. `bun run build` from `landing/` must pass.

Read `landing/src/blog/README.md` before you start and follow it. Never
hand-edit generated per-post HTML or og and title tags. They are generated from
`posts.ts`, so an edit there is overwritten on the next build.

Sitemap: https://orcapods.ai/sitemap.xml

## What we sell, and who reads this

Orca runs AI agents in isolated cloud sandboxes on a schedule, metered to the
micro-dollar, so teams do not need a machine that stays on or hand-rolled safety
rails.

Readers are developers, founders and platform engineers who build agents with
Claude Code, the Agent SDK, Codex, LangGraph or MCP, and need to run them
headless, scheduled and safely.

## Allowed lanes

1. Running AI agents in production: cost, reliability, failure modes, spend caps.
2. Headless and scheduled coding agents: claude -p, cron and launchd, overnight
   runs. This is the highest-value lane.
3. Agent sandboxing and isolation.
4. Agent cost control and metering: token costs, MCP schema overhead, per-run
   billing.
5. Comparisons and alternatives for agent runtimes.

## Forbidden lanes

Generic what-is-AI explainers, prompt-engineering listicles, AI news, model
training or GPU content, consumer chatbot and creator-tool content, crypto or
trading bots, growth-hacking, SEO about SEO. Never name outreach prospects or
customers without written permission.

## Voice

An engineer who runs the platform. First person plural, building in public,
direct and concrete, honest about failures and costs.

The reader is a developer: sandbox, cron, token and API need no explanation.
Orca-specific terms get one plain clause on first use.

Hard style rules:

- No hype, no marketing adjectives, no exclamation marks.
- Contractions always.
- Zero em dashes and en dashes anywhere.
- No sentence under 6 words.
- No run-ons. Split anything chaining three or more clauses. Roughly 30 words is
  the ceiling.
- Mention Orca only where it belongs, 2 to 5 times per piece. No bespoke CTA
  block; the template renders the CTA.

## Claims and numbers

Never invent customers, case studies, usage numbers or benchmarks. Use verified
numbers only, with honest qualifiers. Never promise uptime, an SLA or SOC 2.
Never state pricing from memory. Currency is USD, always.

## Visuals

Background near-black #121212 (hsl 0 0% 7%). Accent coral vermilion #F45434
(hsl 10 90% 58%). One accent element per figure.

No photoreal or AI stock-style images, ever. Figures are live animated React
components, real code, or diagram GIFs. Heroes are typographic stat cards at
1200x750, in the style of landing/public/blog/hero-217kb-5kb.png.

Add anything a glance at the folder would not reveal

If a post has to carry front-matter fields, update an index or registry file, put images in a particular place, or pass a build check, say so. Copy the field names from a real post. This is the difference between a draft you merge and a draft you rewrite.

Save it

You attach it to the scheduled run in the next step, and you can edit it any time afterwards. Edits are picked up on the very next run, with nothing to redeploy.

Orca calls these Prompts in the dashboard and contexts in the concept docs. They are the same object.

Step 5: Set up the scheduled run

The template brought a scheduled run with it, and it arrives stopped so that nothing happens until you are ready.

Open Scheduled Runs

In the sidebar, under Automation, open Scheduled Runs and click the run that came with the template.

Attach your site prompt

On the Context tab, attach the prompt you wrote in Step 4. Without it the run stops at pre-flight and tells you exactly what was missing, rather than guessing at a repository.

Set a time you will actually read drafts

On the Schedule tab, pick the time of day and your timezone. Something early enough that the pull request is waiting when you start work.

Check the dependencies

The Dependencies tab lists everything this run leans on outside its prompt: its skill, its capabilities, and its connected apps. GitHub should be there and healthy before you go further.

Leave it stopped for now. Test first.

Step 6: Test it with Run now

Press Run now

Run now starts one run immediately and does not touch the schedule.

Watch it work

Open the run to follow it live. A full run takes a few minutes: it picks a keyword, reads the current top results, drafts the post, briefs the media agent for figures, then opens the pull request.

Step 7: Check your email and the pull request

Open the email

A healthy run emails you a link to the pull request it opened.

Read the pull request

On GitHub you should find a draft pull request labelled seo-draft on a branch named seo/<slug>, carrying the post, its images, and any index or sitemap updates in the same branch.

Decide

Read it. Merge it if you are happy, and merging is what publishes it. Nothing reaches your site until you do.

Step 8: Start the schedule

Once a test run has produced a pull request you are happy to have received, go back to the scheduled run and press Start. From then on it runs at the time you chose.

When something goes wrong

A blocked run does not guess. It stops, records exactly what was missing, and opens nothing.

What you see	What it means	What to do
The agent reads Needs setup and runs are refused	A required dependency is unresolved	Open the agent and use the Connect button on the banner row that names it
Run stops immediately, mentions missing site facts	No prompt was attached, or it is missing the domain, the repo, or the lanes	Attach your site prompt on the run’s Context tab, and check all three facts are in it
Pull request opened on the wrong repository	The prompt names a different `owner/name`	Fix the repo line in your prompt. The agent only ever writes to the repo you name
Runs are refused and the banner names `composio-dataforseo`	DataForSEO is not connected, or its credentials lapsed	Connect it from the banner row. Check your DataForSEO balance has not run out
No email arrived, but a pull request exists	Delivery succeeded, the notification did not	Check the run’s own output for the link
The draft reads poorly	Usually the prompt rather than the agent	Tighten the voice and style rules in your site prompt, then run again

Still stuck? Email support@okik.io with the name of the scheduled run and roughly when it ran, and we will look at it.

Making it better week by week

The first few drafts will need work, and that is normal rather than a sign the setup is broken. When a draft disappoints you, do not stop at fixing the draft. Edit the pull request by hand if you like, or hand it to another agent to tidy up, and then go back and change whatever let it through:

Something specific to your site, like voice, subjects, file conventions or length, belongs in your prompt.
Something about the method, like the keyword rules or the writing rules, belongs in the seo-engine skill, which is plain markdown you can open and edit.
Something about how an agent behaves overall belongs in the agent itself, including which model it runs on.

Every one of those is a text edit in Orca, and every fix holds for every run afterwards. Most of the effort lands in the first week. After that it settles into a short read each morning.

Shared templates

How copying, renaming and updating templates works.

Connected apps

Connecting GitHub, DataForSEO and 250+ other apps.

Prompts

Facts about the work, attached to a run.

Source

The agents, the skills and the method come from the SEO Agent Pack by Nicolas Lecocq, MIT licensed. Read it there to see how the method works underneath, or to adapt it for your own kit.

Get Started

Core Concepts

Guides

Runbooks

SEO helper: an agent that drafts your blog posts

Before you start

A GitHub repository

A DataForSEO account

Step 1: Open the template and copy it

Step 2: Connect GitHub

Step 3: Connect DataForSEO

Step 4: Write your site prompt

Step 5: Set up the scheduled run

Step 6: Test it with Run now

Step 7: Check your email and the pull request

Step 8: Start the schedule

When something goes wrong

Making it better week by week

Shared templates

Connected apps

Prompts

Source

​Before you start

A GitHub repository

A DataForSEO account

​Step 1: Open the template and copy it

​Step 2: Connect GitHub

​Step 3: Connect DataForSEO

​Step 4: Write your site prompt

​Step 5: Set up the scheduled run

​Step 6: Test it with Run now

​Step 7: Check your email and the pull request

​Step 8: Start the schedule

​When something goes wrong

​Making it better week by week

​Related

Shared templates

Connected apps

Prompts

​Source

Before you start

Step 1: Open the template and copy it

Step 2: Connect GitHub

Step 3: Connect DataForSEO

Step 4: Write your site prompt

Step 5: Set up the scheduled run

Step 6: Test it with Run now

Step 7: Check your email and the pull request

Step 8: Start the schedule

When something goes wrong

Making it better week by week

Related

Source