Seven prototypes.

How Orkonerei FM built OrkoBot

How they built it

1. 01
   Get trained first

   The build followed hands-on AI training; the builder didn't have the knowledge to create an agent before. Two skills matter most: creating a custom agent, and detailed prompt-writing.

2. 02
   Define the problem as a team

   The whole team discussed what they needed.

3. 03
   Choose the platform

   ChatGPT, on a paid subscription.

4. 04
   Create the agent and ground it in your own documents

   They fed it the station's mission, vision and task descriptions so it can understand their mission and vision and the tasks needed.

5. 05
   Test, then open to everyone

   Tested first, then shared with all team members.

6. 06
   Fix it into the workflow with human checks

   The documented chain starts with a story idea, refined with OrkoBot for topic, angle, programme plan, interviewees and format. This is followed by field reporting, AI-supported script drafting, human editing and fact-checking, and then broadcast, website or social media. The review rule is written into the station's policy. The journalist reviews first, then a role-specific checker.

Time and money

Roughly three months from training to team use. Costs cover the AI and internet subscriptions. The three AI licences together cost USD 1,431 for 12 months.

What to expect

The challenge is prompting, not the technology itself. Results at Orkonerei show three to five posts a day and marked follower growth over the project, with the newsroom's own careful caveat that AI workflows "combined with newsroom efforts" contributed, alongside a YouTube channel launch.

Check before you copy

Which languages the agent must work in, and how well it performs in yours.

How Lake Region Bulletin built its fact-checking agent

How they built it

1. 01
   Get trained

   The team describes itself as initially wary of AI, with no strong background in it. Hands-on training helped them learn they could "code using words," configuring an agent purely through natural-language instructions.

2. 02
   Define one specific task

   Make the fact-checking desk's manual work faster. Most mis/disinformation they see is recycled, such as a Nigerian photo passed off as a Kenyan event, so the task was tracing where content first appeared.

3. 03
   Choose a custom agent over building from nothing

   Building a new tool was judged too technical and too expensive (cloud hosting, maintenance, monitoring) for their scale. A configured agent on a paid consumer subscription was affordable and fast.

4. 04
   Configure it in three steps

   Explain to the agent who the newsroom is and what it wants to achieve. Feed it with instructions on how to fact-check, plus reference resources. Upload several of your own published fact-checks so it learns the house format.

5. 05
   Run it with human supervision

   The agent traces provenance; the editors debunk; findings are fed back in and the tool drafts a fact-check story in house format. A human reviews before anything is published. No AI-caused mistake has ever gone to press.

6. 06
   Iterate the prompt, not the tool

   When outputs failed the "human test," they added more resources and more specific instructions to narrow what it searches for.

7. 07
   Restrict access until you trust it

   Only the two editors use it, "to regulate its access for proper management," with wider onboarding only "once convinced it is safe."

Time and money

The main cost is the paid subscription, around USD 25–30 a month, found sufficient after testing tiers from the smallest upward with unlimited usage. A secondary cost is supporting staff home Wi-Fi, since the team works remotely.

What to expect

Persistent hallucinations (in one case the tool asserted Barack Obama is the current US president), which is why the human stays in the loop. Some websites block AI access, so information is sometimes unreachable. Imprecise prompts send the tool "fishing" wrong information. Free tiers are limiting, and the tools keep changing. The desk now averages two fact-check stories a week, plus general verification of figures in other reporting.

Check before you copy

You must supply your own house-format fact-checks, and your own local threat picture. LRB's agent is tuned to recycled foreign content re-captioned for Kenya; yours needs tuning to whatever circulates where you are.

How Chequeado built Nunchi, its audience-comment agent

How they built it

1. 01
   Start from a clear problem, with numbers

   Chequeado watched metrics but was not reading its comments of roughly 10 Instagram posts a week averaging a thousand comments each, around 10,000 a week from Instagram alone.

2. 02
   Buy the expertise, keep the capability

   An external n8n specialist set up the initial workflow, deliberately skipping the learning curve while training the team, so they can support the tool themselves.

3. 03
   Check what is free before you pay

   They expected to pay for an n8n licence, then discovered it is open source and free if self-hosted, which reduced operating costs and freed that budget.

4. 04
   Narrow the scope until it works

   They first tried to extract many facets per comment and stalled. Cutting to far fewer dimensions unblocked progress. Over-scoping the categorisation is the predictable trap.

5. 05
   Involve the newsroom and show the raw material

   Three or four one-hour UX sessions with the wider team. The key decision is to include the real comments inside each Slack alert, so the team can check the AI against the source and trust the system. Their own hindsight lesson is to define the right moments to involve the newsroom, balancing showing a prototype too early against too late for useful feedback.

6. 06
   Calibrate checking to risk

   No one is assigned to verify every categorisation; the team keeps an eye on it and reads the raw comments when editorial flags a bad output. That lighter validation is acceptable because the output is internal-only. Audience-facing output would require stricter processes.

Time and money

API usage runs roughly USD 50–100 a month at current volume, expected to rise if other platforms are added. The n8n licence costs nothing self-hosted. Build time was around 10 to 15 hours a week.

What to expect

Spanish-language comments worked far better than initial thoughts. One or two pieces have been published based on insights the comments surfaced. The roadmap includes more platforms, new use cases on the accumulated database, adoption beyond the newsroom, and possibly opening the tool to other newsrooms as a sustainability model.

Check before you copy

Your comment volume and platform mix set both the value and the cost. The categorisation they settled on focused on sentiment in six registers, intent, and signals such as support, criticism of the outlet, and conspiracy theories.

How Radio Domus is building Safeguard

Safeguard is a prototype in refinement, not a finished product; this guide covers the path so far.

How they built it

1. 01
   Start from a local failure of global tools

   Facebook flagged an audience comment calling a show "a bomb" (meaning excellent) as terrorism. Meanwhile genuinely abusive Swahili and Sheng terms pass unflagged. That gap, in both directions, is the problem Safeguard exists to close.

2. 02
   Build a keyword library from how your community actually speaks

   Over 500 terms across English and Kiswahili covering hate speech, ethnic discrimination, incitement, profanity and election manipulation, fed into the AI agent to contextualise moderation. Expansion planned ahead of Kenya's 2027 election. A starter dataset exists and is shareable.

3. 03
   Choose build-and-integrate, not build-or-buy

   Market tools alone would not work for a community station, so they're building Safeguard with a paid Claude subscription while integrating third-party tools (NapoleonCat, the Meta dashboard) and feeding local knowledge into them.

4. 04
   Get help for what you cannot do, but learn the backend

   The second phase onboarded two external organisations for the front end, judged too complex to do alone, while the team deliberately learned the backend so the skills stay in-house.

5. 05
   Put the human loop in cheap, common tools

   The human-in-the-loop mechanism lives in Slack, not inside Safeguard, where journalists submit drafts, an editor reviews for accuracy and balance, revisions follow, approval precedes publication. The tool alerts whoever is online for post control. A human makes the final call on everything.

6. 06
   Write the ethics before the automation

   The operating model revolves around AI suggesting, human reviewing, and decision logging. Core protections include data minimisation, role-based access to audience data, no sensitive data in open AI tools, and review of false positives and negatives. The team says building these guardrails first deliberately slowed automation.

7. 07
   Scope to your highest-volume platforms and test with a lean team

   Facebook and YouTube first; Instagram, X and TikTok deferred. Interns support testing and iteration because everyone multitasks.

Time and money

Deliberately low cost, but the stack is wide. Subscriptions to Claude, ChatGPT, Slack, NapoleonCat and Trello, a Microsoft 365 trial, and staff training.

What to expect

They underestimated the technical expertise required. One AI subscription alone left the social media side uncovered, so complementary tools had to be factored in. The dashboard is unfinished and no impact analytics exist yet. Early observed gains are at the workflow level, with structured identification of harmful content, quicker review, clearer escalation, more consistent decisions.

Check before you copy

The keyword library does not transfer: you must run your own gathering exercise in your own community's languages. Expect to need external technical help for the front and backend, and budget for a small stack of tools, not one subscription.

How La Diaria built Núbel, its weather-alert article generator

How they built it

1. 01
   Start from a task nobody wants

   The web team had to turn every red, orange or yellow alert into an article manually, essentially reprinting a press release with "no soul" and no added journalistic value. A bad-weather stretch meant doing it every morning for two weeks. The team asked for help, which made adoption immediate later.

2. 02
   Use a free prototyping environment

   The tool was built in Google AI Studio, an environment for building AI prototypes, in a roughly two-week sprint. The only additional work was wrapping it in a web interface journalists could use.

3. 03
   Teach it your style with your own articles

   Núbel learned the house style from the paper's own past forecast articles, so its output reads like La Diaria, not like a press release.

4. 04
   Keep a human on the figures

   A journalist verifies every number before publication, including wind speeds, rainfall amounts, the departments affected. The tool drafts; it does not publish.

5. 05
   Name an owner before the sprint ends

   The tool cut the task from about 20 minutes to 10, and adoption was immediate. By mid-2026 it was nonetheless not running. The staff member who drove it changed sections, and no routine existed to check whether tools still work.

Time and money

A two-week sprint by the participants, a free build environment, no internal IT time.

What to expect

Fast results on any structured official bulletin that arrives in a consistent format. The risk is not the build but the afterlife. Without an institutional owner, even a successful tool quietly stops.

Check before you copy

You need a bulletin with a stable format, your own past articles to teach the style, and a named person to verify figures on every output.

How La Diaria built TIA, its one-button title assistant

How they built it

1. 01
   Find the friction point

   Journalists were not writing meta titles. It was an extra step nobody took at the end of a long writing day, so articles were not appearing in search or on Google Discover. With subscriptions over 80% of income, reach beyond subscribers matters commercially.

2. 02
   Fit the tool into the workflow, not beside it

   TIA was deliberately built as a tiny step inside the existing publishing flow rather than a standalone product. One button, three suggestions, all editable. Human review is the design, not an add-on.

3. 03
   Calibrate with house examples

   The tool was calibrated with the paper's own articles paired with correct meta titles.

4. 04
   Give it to the whole IT team

   The internal IT team did the technical work across roughly three months on top of normal duties, its first AI project.

5. 05
   Control access, then let demand spread it

   Access is permission-controlled. Adoption grew by word of mouth from three initial editors to all web editors and journalists, with colleagues requesting access, the newsroom's strongest proof of success.

Time and money

A monthly AI subscription, plus roughly three months of the internal IT team's time.

What to expect

Strongest on short, fast-turnover content such as breaking news and sports, but poor on long articles the design never considered. Journalists also fed the suggestions back into actual headlines, a use nobody planned. Like Núbel, the tool later stopped working for lack of an owner.

Check before you copy

You need a CMS you can modify, some internal IT capacity, and house-style examples from your own pages. Design it for your long-form content from the start.

How Anfibia built its Slack chatbot for visual content

How they built it

1. 01
   Run a needs assessment first

   Anfibia interviewed each area's leaders about what they needed. The problem that surfaced: every branch of a multi-arm organisation needs social visuals, and all of it landed on an art department of two to three people.

2. 02
   Write a brief, then hire competitively

   The team wrote a specification, posted a hiring call for developers on social media, interviewed three software engineers and selected one it had worked with before.

3. 03
   Build iteratively, then have the developer train the users

   The build took roughly two to three months with feedback from the newsroom, followed by user testing and two short developer-led training sessions for area leaders.

4. 04
   Pilot on the lowest-risk content only

   The chatbot handles the easiest requests and is deliberately not used for imagery attached to journalistic articles. Anfibia learned that boundary the hard way. An AI-generated gallery once labelled the Falklands/Malvinas as British on a map, published without supervision, and the audience's reaction was furious.

5. 05
   Keep the approval loop mandatory

   Every output travels with the requester's brief to the art director, who must approve before anything is used. Faulty outputs are simply discarded.

6. 06
   What they would do differently

   Involve the real end users from day one. The project was designed from the art department's perspective, but the daily users turned out to be the social media team, who joined mid-project. The feedback that mattered most arrived late.

Time and money

The external developer's fee for the two-to-three-month build, an n8n licence, two or three ChatGPT seats, plus existing design-software subscriptions.

What to expect

An estimated five to seven pieces of content a week that would otherwise have gone to the art department, with the tool still in testing. Not everything generated gets used.

Check before you copy

Recruitment is the real variable. Anfibia found its developer through a tech reporter's personal network. Make sure to budget for hiring and training. Check you have enough routine visual demand to justify the build; Anfibia's many branches generate unusually heavy flyer traffic.