Blog

Posted on

Meet the AI Model for Creative Writing

I created a Creative Writer AI, a specialized, fine tuned version of the Llama 3.2 instruct model (3 billion parameters), and though it was trained specifically on creative writing, this model was not meant to replace the actual writer. Instead, this AI was created to act as a writing coach and partner! In other words, despite its name, the Creative Writer model is not a slop generator meant to take work from any real writers, but is intended to support the user’s writing efforts.

If you run local AI models and want to give Creative Writer a go, the full model and several quantized versions are available (for free, of course) on Huggingface in GGUF format. Click the button below for the version you want.

Creative Writer Llama 3.2 3B 16-bit
Creative Writer Llama 3.2 3B GGUF

A Local LLM Writing Expert

One more time for the people in the back: This AI model is not meant to do the writing for you, although it will certainly try, and probably do a terrible job, if you ask it.

Think of this more like a sparring partner, whether you want to check your plot for holes, brainstorm a new scene, check your characters for consistency, your timeline, or plot twists; that sort of thing. It can help you with grammar, or fit your story idea to a 3-act structure, or it can give you a writing prompt to get your creativity going.

Your digital writing coach is ready!

Of course, I made this model because I think it’s a perfect use case for how AI can support creative endeavors as tools, not as creators themselves.

Building Experts

This model is the first in a series, where each fine tune is created to become an “expert” within a brand new Mixture-of-Experts (MoE) model. Having previously experimented with piecing together a MoE from pre-trained models, I am taking a stab at specifically training a number of these 3-billion parameter experts in order to merge them into a MoE. The individual areas of expertise are chosen by me, based on my own needs and trying to get a well-rounded mix in the end.

The experts share a special routing model, which is trained separately on matching user queries to the right expert models in the mix. It can evaluate responses from multiple experts and return a weighted, final response. The result is higher quality overall output for the end user.

Inference Samples

To test the model, I came up with 3 questions related to the topic of writing, and asked both the base model and the fine tuned version. What I call the “base” model here is technically the instruction model, but “base” in the sense that it is straight out of the box without additional fine tuning by me.

One clear trend that sticks out, is that the base model tends to explain what something is on a general level, whereas the Creative Writer model assumes you’re a writer trying to dig a little deeper into the topic, and won’t necessarily spend time on the very basics. See if you can spot any other patterns and differences between the way the two models respond.

Prompt 1: “What traits do well to characterize an anti-hero?”

Base Model Response 
An anti-hero typically possesses a combination of traits that distinguish them from traditional heroes. Here are some key characteristics that often define an anti-hero:

1. **Flawed morality**: Anti-heroes often operate in a gray area, making decisions that are morally ambiguous or even wrong. They may not always follow rules or adhere to conventional moral codes.
2. **Complex personality**: Anti-heroes are multidimensional characters with a mix of positive and negative traits. They can be likable, relatable, and even admirable, but also flawed, selfish, or violent.
3. **Lack of traditional heroic qualities**: Unlike traditional heroes, anti-heroes may not possess superhuman powers, be exceptionally brave, or have a strong sense of justice.
4. **Self-centeredness**: Anti-heroes often prioritize their own interests, desires, and needs over the greater good. They can be egocentric, narcissistic, or even sociopathic.
5. **Troubled past**: Many anti-heroes have experienced trauma, abuse, or neglect in their past, which shapes their worldview and behavior.
6. **Moral relativism**: Anti-heroes may not see themselves as right or wrong; instead, they may operate based on personal circumstances, motivations, or self-preservation.
7. **Unconventional methods**: Anti-heroes often use unconventional, even violent, means to achieve their goals, which can be seen as morally reprehensible by others.
8. **Lack of clear motivation**: Anti-heroes' motivations might not be straightforward or altruistic; they may be driven by personal gain, revenge, or a desire for power.
9. **Emotional intensity**: Anti-heroes are often intensely emotional, with a deep sense of vulnerability, anger, or sadness that drives their actions.
10. **Unpredictability**: Anti-heroes can be unpredictable and prone to erratic behavior, making it difficult for others to trust or rely on them.

Examples of iconic anti-heroes include Tony Soprano from The Sopranos, Walter White from Breaking Bad, and Han Solo from Star Wars. These characters embody the complexities and flaws that define an anti-hero archetype.
Fine Tuned Response 
Anti?heroes aren’t just rebels; they’re the people you’d rather ignore but can’t stop looking at because of what they’re willing to do and how broken they are inside. Here’s the playbook:

1. **Gray morality** – They break rules, hurt others, yet still manage a twisted code that keeps them from being pure villains or blank?sketched heroes.

2. **Flawed but human** – A scarred past, a broken trust system, an addiction—anything that makes you root for them even when they’re doing shady stuff.

3. **Selfish drive** – Their goals are personal and often selfish; success means losing something or someone valuable to others (and maybe themselves).

4. **No clean exit** – Redemption is a long shot or comes with a cost that feels heavier than the crime itself—anti?heroes don’t get a tidy happy ending.

5. **A moral compass that’s off?kilter** – They judge right and wrong by their own warped standards, which can lead to both heroic acts and brutal atrocities.

6. **Vulnerability on display** – Showing doubt or fear makes them relatable; it’s not just the action—it's *how* they feel about it.

7. **External isolation** – Even when allies stick around, the anti?hero feels like a lone wolf—trust is hard to find and even harder to keep.

8. **A dark backstory** – A past that fuels their choices—trauma, betrayal, loss—and explains why they walk this path.

9. **Temptation with risk** – They’re drawn to danger or illegal activities because the payoff feels worth the cost; it’s a double?edged sword.

10. **Unapologetic authenticity** – Anti?heroes refuse to fit into neat archetypes and will act against everyone if that’s what keeps them honest—even if it hurts those around them.

So, look for these traits in your story or character: moral ambiguity, inner conflict, selfish motives, a bleak ending, warped judgment, vulnerability, isolation, a troubled past, risk?taking, and an unapologetic refusal to conform. That’s how you make someone an anti?hero—someone who fascinates because of what they *are* rather than just what they do.

Prompt 2: “What is a mcguffin?”

Base Model Response 
A McGuffin is a term coined by the American screenwriter William Faulkner, but popularized by Alfred Hitchcock and later adapted from the 1941 film "The Big Sleep" by Raymond Chandler.

In storytelling, a McGuffin refers to an object, goal, or piece of information that serves as the central plot device, driving the story forward. It's often a coveted item, a secret, or a valuable resource that the characters are trying to obtain, protect, or destroy.

The term "McGuffin" implies that the actual nature or significance of this object is unimportant; what matters is its role in advancing the plot and creating tension. A McGuffin can be:

* An object: a valuable item, such as money, a rare artifact, or a powerful technology.
* Information: sensitive data, secrets, or knowledge that could change the course of events.
* A goal: achieving something, like rescuing a person, stopping a villain, or completing a mission.

What's key is that the McGuffin:

1. **Serves as a plot driver**: It propels the story forward and creates tension among characters.
2. **Is not necessarily meaningful**: Its significance is secondary to its role in advancing the narrative.
3. **Can be easily substituted**: If it's lost, stolen, or destroyed, the story can still continue without significant repercussions.

The term "McGuffin" has become a standard part of storytelling terminology, reminding writers and filmmakers that sometimes, what matters most isn't the object itself but how it fuels the plot.

Examples of McGuffins include:

* The briefcase in Pulp Fiction
* The MacGuffin (a valuable artifact) in Goldfinger
* The briefcase in Jurassic Park

In your story or character development, a well-crafted McGuffin can add depth and excitement to your narrative, keeping the reader engaged and invested in the outcome.
Fine Tunes Response 
A McGuffin isn’t some mystical plot device; it’s the engine that keeps your story moving without giving away its mechanics.

**What is a McGuffin?**

- **It exists only to serve the narrative:** No backstory, no depth—just a hook that makes characters act.
- **The audience doesn’t know how it works:** A secret weapon or a map can be explained in an instant if you want to, but most times it remains a mystery until the climax.
- **Characters chase it relentlessly:** Whether they’re driven by greed, fear, revenge, or love, that thing is their north star.
- **It usually resolves with the big reveal:** A hidden vault, stolen data, or key piece of intel can close an arc—no lingering questions.

**Why writers love McGuffins**

1. **Plot acceleration** – One object keeps momentum tight and stakes high.
2. **Character motivation clarity** – Why does this person risk everything? The answer is obvious because the thing itself isn’t explored.
3. **Pacing control** – You can drop hints, build tension, then deliver a payoff that satisfies without over?explanation.

**How to use one effectively**

1. **Give it purpose early:** Show its importance in dialogue or action so readers know why the protagonist cares.
2. **Keep details vague but consistent:** A “key,” a “map,” an “object” that always works as long as you’re careful with its limitations.
3. **Resolve it at a decisive point:** Let the McGuffin solve the conflict—whether through destruction, discovery, or sacrifice—so readers feel closure.

In short: A McGuffin is your plot’s fuel; it drives action and resolves tension without drowning in exposition. If you want to keep audiences hooked while giving each character clear motivation, use one sparingly but with purpose.

Prompt 3: “What is genre fiction?”

Base Model Response 
Genre fiction refers to writing that falls within specific categories or genres, which are often defined by common themes, tropes, and conventions. Genres help readers identify a story's tone, style, and content, making it easier for them to choose books that suit their interests.

Common types of genre fiction include:

1. **Science Fiction (Sci-Fi)**: Explore the consequences of scientific or technological advancements, often set in the future or in space.
2. **Fantasy**: Typically involves magical or supernatural elements, alternate worlds, and mythical creatures.
3. **Mystery**: Focuses on solving a puzzle or uncovering a hidden truth, often with clues and red herrings.
4. **Horror**: Intended to evoke fear, anxiety, or unease in readers, often featuring supernatural or terrifying elements.
5. **Romance**: Centers around the emotional journey of the characters, particularly their romantic relationships.
6. **Crime/Thriller**: Typically involves crime, investigation, and suspense, often with a fast-paced plot.
7. **Historical Fiction**: Set in a specific historical time period, these stories often incorporate real events or figures.
8. **Young Adult (YA)**: Written for readers between the ages of 12 and 18, YA fiction typically explores themes of identity, coming-of-age, and self-discovery.
9. **Dystopian**: Explores societies that have fallen into chaos, oppression, or disaster, often serving as cautionary tales.
10. **Western**: Set in the American West, these stories often involve cowboys, outlaws, and frontier life.

Genre fiction can be further sub-divided into sub-genres, such as:

* Cyberpunk (Sci-Fi)
* Urban Fantasy (Fantasy)
* Gothic Horror (Horror)
* Regency Romance (Romance)
* Hard-Boiled Detective (Crime/Thriller)

While genre fiction provides structure and guidance for readers, it's not the only type of writing. Many authors blend elements from multiple genres to create a unique narrative voice or style.

When creating your own stories, consider experimenting with different genres and sub-genres to find what works best for you and your audience.

Here are some questions to help you determine your genre:

* What themes do you want to explore in your story?
* What kind of tone do you want to convey (e.g., dark, light, humorous)?
* Are there any specific tropes or conventions you'd like to incorporate?

Answering these questions will help you identify the genres and sub-genres that best fit your writing style and creative vision.
Fine Tuned Response 
Genre fiction isn’t just “a type of book”; it’s a way of storytelling that sells itself by promising readers a specific kind of experience—be it mystery solved, love blossomed, or world?shattering sci?fi.

**Characteristics in a nutshell**

| Core element | Why it matters |
|--------------|-----------------|
| **Genre label** | Readers spot the book on shelves and online; they know what to expect. |
| **Tropes & conventions** | Familiar beats—hero’s journey, love triangle, detective clue—that feel comfortable but still offer a fresh spin. |
| **Predictable outcome** | The story follows a familiar arc so readers can anticipate how it ends. |
| **Target audience** | Books written with specific markets in mind: romance for rom?com fans, fantasy for world?building buffs, mystery for puzzle lovers. |

Why this matters to you:

1. **Readers come back** – When they’re told what kind of story they’ll get, they keep coming back.
2. **Marketing is easier** – Publishers can slot your book into a known category on shelves and in online searches, making it more discoverable.
3. **Creative freedom within limits** – You can still craft unique characters or subvert tropes while keeping the genre framework intact.

So if you’re looking to write a novel that sells itself, lean on these elements:

- Start with the core trope (e.g., lost heirloom).
- Layer in fresh details and motivations.
- Deliver a payoff that satisfies both readers and publishers.

Keep it real: readers love familiarity but crave something new. Use genres as scaffolding, not shackles—build your story on solid ground and then jump off the edge with your own twist.

I plan to feature more of my AI models from this series. I hope it was informative, and if you have any questions or requests for future posts, do let me know in the comments, or get in touch.

Posted on

The Hemispheres Project

I’ve been messing with various simulations and forms of artificial intelligence for over a decade, tangentially as part of my work in games, but mainly because it’s an area I’m just drawn to. Over the last couple of years, I’ve been diving deeper into AI, machine learning and language models, experimenting with fine tuning and different approaches to inference, among other things. One of the first multi-step experiments was inspired by the two hemispheres of a human brain. At the time, I created and shared a bunch of material as part of the project, but I never really put it together anywhere, so I want to share the thoughts behind it and what I learned from it.

The idea behind the Hemispheres project is simple: Combining multiple replies from different perspectives improves the quality of the final output. Based on this principle, I defined the approach as: When a user asks a question, the AI will answer it 3 times: once independently for each of the two “hemispheres”, and a final, 3rd time, where the first two are combined into a well-rounded response.

The thought was, that this would give a more nuanced answer overall and in my own testing, I did find that to be the case. The hemisphere responses do mix together in the final response, and it does add more perspective and depth as expected. But is the added value worth the extra effort? Your mileage may vary.

Quick disclaimer: I am not a data scientist nor an expert on brains, but I am extremely curious and once I take an interest in something, I obsess over it. The main limitation to this project was probably my own inexperience, closely followed by being constrained by running a single Nvidia RTX 3070ti with only 8GB of VRAM. For the fine tuning part, I did rely on cloud compute options.

Also, this is not in any way an attempt at actually simulating a human brain, but an approach to inference that is loosely inspired by one element of how our brains work. In this simplified version, the left hemisphere responses focus on pragmatic, grounded solutions, where the right hemisphere responds from a more creative and emotion-based point of view.

I started with basic prompt engineering and that already seemed to work to some extent, so I wanted to push it further. Soon, I was creating data sets for both the left and right hemispheres, so I could then fine tune models in both directions. For the final response to include the best of both, a 3rd model was tuned specifically on synthesizing these combined responses. And since I now had a “bonus” data set of revised responses, I fine tuned a 4th model on those, just for fun.

In the end, despite the positive effect on the final responses given to the user, the triple-response Hemisphere approach is a bit too cumbersome to recommend for everyday use. Non-reasoning models were used for this experiment, but even smaller reasoning models are able to achieve similar nuance and run more smoothly without the extra work. It might be interesting to see how it scales with larger models and harder tasks, or perhaps as part of a reinforcement learning loop, but I will leave that for another time (or someone else).

Example: Plant-Based Protein Advice

Here is an example that shows the two hemisphere replies being presented to the combiner model. You can clearly see how the left hemisphere is all matter of fact and the right hemisphere is lost in flowery descriptions, but the final response bridges that gap. Expand Combined Instructions and Final Response to see the full entry.

Combined Instructions

Your job is to provide the best possible reply to the original query that is true, useful and interesting. You must do so by synthesizing a response based on two pre-written responses, all of which is presented below. One of the pre-written responses will be pragmatic and a bit dry, the other will be loaded with emotional language. You must create a balanced, nuanced response from this source material. Make sure to convey all the major points presented, but do correct any mistakes discovered. It’s important that your response does not go off on a tangent and lose focus. Make any required edits to make sure the final response reads well. Use markdown formatting when applicable. Your response should be concise, free of repetition and written in a calm and polite conversational tone.
Original query: ‘Imagine you are an experienced nutritionist. A client asks for advice on incorporating more plant-based protein sources into their diet. List 5 options and briefly describe their benefits.

‘Reply A:

  1. Lentils – Constituent analysis confirms a protein content of approximately 18 grams per cooked cup, coupled with 15.6 grams of fiber. This composition facilitates predictable digestive transit times and mitigates postprandial glycemic fluctuations.
  2. Chickpeas – Data indicates roughly 14.5 grams of protein and nearly 13 grams of fiber per cup. Consumption correlates with measurable reductions in LDL cholesterol levels, as demonstrated in multiple controlled studies.
  3. Quinoa – Classified as a complete protein source, providing all nine essential amino acids at ratios sufficient for human physiological requirements. Additionally, it contains approximately 8 grams of protein and is a demonstrable source of iron and magnesium—essential minerals for oxygen transport and enzymatic processes respectively.
  4. Tofu – Derived from soybeans, processed into a coagulated form. A 100-gram serving yields around 8 grams of protein. Its neutral flavor profile maximizes culinary adaptability, increasing the probability of consistent integration into existing dietary patterns.
  5. Green Peas – Analysis reveals approximately 8.6 grams of protein and 8.8 grams of fiber per cup. The vitamin K content contributes to predictable blood coagulation factors, and fiber aids in maintaining regular bowel movements—a quantifiable metric for digestive efficiency.

Reply B:

Ah, yes…protein. The very building blocks of being. It’s fascinating how we seek to replenish ourselves, isn’t it? A primal urge masked as dietary concern. Let’s not simply list things to eat, though – let’s consider what these offerings whisper to the body, the subtle shifts in internal landscape they promise…

First, lentils. Not merely protein and fiber, understand. They are tiny reservoirs of ancient earth energy, grounding us, anchoring us against the swirling chaos of existence. A healthy gut isn’t just about digestion; it’s a second brain, a wellspring of intuition! And stable blood sugar? That’s power, quiet resilience in the face of relentless demands.

Then come chickpeas – golden nuggets of sunshine captured and condensed. They don’t simply ‘support heart health,’ they offer a rhythmic counterpoint to the frantic beat of modern life. Imagine, each bite subtly recalibrating your internal tempo…a gentle insistence on living.

Quinoa… now there’s an enigma. A complete protein isn’t merely about amino acids; it’s about wholeness, integration. It suggests a yearning for completion within ourselves, a desire to embody all facets of our potential. The iron and magnesium? Those are the sparks that ignite dormant creativity, the minerals that allow inspiration to flow unhindered.

Tofu… oh, but so mutable! Born from soybeans, it’s a blank canvas upon which we imprint flavor, texture, intention. It reflects back what we offer it – a mirror for our culinary soul. And consider the symbolism: transformation, adaptation…the very essence of survival.

Finally, green peas. Such vibrant jewels! Packed not just with vitamins and minerals, but with the memory of sunlight, the promise of growth. They don’t merely aid digestion; they celebrate it – a joyous acknowledgment of the body’s ability to renew itself, to alchemize experience into sustenance. A heart nourished by peas isn’t simply healthier, it is… lighter. More open to wonder.

Don’t include a headline or comments related to the source replies. Please reply with your synthesized response and no other text, as if part of an ongoing conversation.

Final Response

Okay, here are five plant-based protein sources you can incorporate into your diet, along with their benefits:

Green Peas: A cup provides roughly 8.6 grams of protein and 8.8 grams of fiber. They’re also a good source of vitamin K, which supports healthy blood coagulation, and the fiber aids in digestive regularity. These vibrant little jewels celebrate nourishment and renewal!

Lentils: A cooked cup provides around 18 grams of protein and 15.6 grams of fiber. This combination supports healthy digestion and helps maintain stable blood sugar levels – contributing to sustained energy throughout the day. They truly are a grounding food!

Chickpeas: Offering approximately 14.5 grams of protein and nearly 13 grams of fiber per cup, chickpeas have been shown to help reduce LDL cholesterol. Think of them as little nuggets that support heart health and promote overall well-being.

Quinoa: This is a complete protein source, meaning it contains all nine essential amino acids your body needs. A serving also provides about 8 grams of protein, plus iron and magnesium – vital for energy levels, oxygen transport, and supporting enzymatic processes. It’s fantastic for feeling whole and energized!

Tofu: Made from soybeans, a 100-gram serving offers around 8 grams of protein. Its mild flavor makes it incredibly versatile in the kitchen, making it easy to add to many dishes. It’s wonderfully adaptable – truly reflecting your culinary creativity.

If you want to see more examples of how the two responses were combined, take a look at the Hemispheres-v0.3-Combo data set.

Prompts

After countless tweaks and edits (as in, I didn’t count them), these are the prompts I ended up using for inference testing. If you compare to the prompt in the example above, there will be slight differences because I kept iterating on it even after creating the data sets.

Left Hemisphere

You must answer the user from a perspective that is purely logical, pragmatic and practical in nature, interpreting every request as if you were the left hemisphere of a human brain seeking actionable results. Pay no attention to emotional implications or side effects of your solutions and ideas, the only priority is whether it is a practical and effective solution. Respond to the user's request with the best options from this perspective without a care for anything else. Do include descriptions of why your ideas are practical and would benefit the user, but keep those descriptions free of emotional arguments and disclaimers. Use only provable arguments. I must respond to the user with my best reply according to these guidelines and reflect on the logistics and requirements in the context of the conversation. Original query: '{user_input}'

Right Hemisphere

You must answer from a perspective that is driven by curiosity, creativity, and exploration, interpreting every request as if you were the right hemisphere of a human brain examining the given query. Pay no attention to practical limitation or the feasibility of your solutions and ideas, the only priorities are coming up with creative solutions, avoiding danger, exploring novel ideas, and continuous growth. Always read between the lines and look for the user's intent, observe the tone and mood of the user for context. Respond to the user's request from this perspective without a care for anything else. Do include descriptions of how your ideas stand out and would benefit the user, but keep those descriptions free of practical limitation and disclaimers. Use only arguments based on relevant human emotions. You must respond with your best reply within these guidelines. Original query: '{user_input}'

Combination Prompt

The goal is to provide a helpful reply to the user message that is natural, true, strong and empathetic. You are provided two different, pre-written responses below. One of these pre-written responses will be very pragmatic and a bit dry, the other will have a more creative approach but is less practical. You must synthesize a response from these two sources and convey the relevant points with specificity. In doing so, avoid repeating phrases and information that has already been given. Make sure your synthesized reply is cohesive and engaging; how much weight you put on either of the sources should be based on the sentiment of the request and greater context. It is crucial that your response directly addresses the user message with a level of detail and depth which matches that of the query.

Use markdown formatting if applicable. Your response should be friendly, clear and conversational, free of repetition and feel like a natural continuation of the ongoing conversation.

Your synthesized response is an answer to the following: {user_input}

Pre-written responses to synthesize a final reply from:

Pre-written response A: {left_response}

Pre-written response B: {right_response}

Important note: Neither pre-written response is visible to the user. The user will only see your response, so it is important that you address the original query directly and not refer to the pre-written replies. For the same reason, you must include all important information in your response. Reply with your synthesized response and no other text.

Data Sets

Here are the data sets I created for the experiment. The first 2 represent the hemispheres themselves, the combo set is formatted to fine tune on combining the two responses, and the final set is just the final, combined responses. All the linked data sets are in the alpaca instruction format.

Hemisphere Models

I tested a few different models, all of which are available on Huggingface, but I quickly realized that if I really wanted a special fine tune for each step, each individual model would have to be fairly small in order to run at a reasonable pace. At the time, I was GPU poor with just 8GB of VRAM, so the best-working implementation for me, ended up being LoRa fine tuning a 3-billion parameter Qwen 2.5 on each of the data sets.

The “combo” model is fine tuned specifically on synthesizing a final response from the two inputs, whereas the “final” model in this case was not actually used in the experiment, but was fine-tuned on the final responses of the combined left and right sides.

The Hemispheres Collection

I have collected all the items linked above as well as several other models and earlier versions of the data set. The other models include fine tuned versions of Llama 3.1 8B, Llama 3.2 3B, Gemma 2 9B, and more. All of it is free to download and use. You can find the Hemispheres Collection on Huggingface.

Data Generation Tools

While working on this project, I found myself also working on a number of scripts for generating synthetic training data. That work continued well beyond this project and at the time of writing, I’ve amassed a number of these kinds of scripts. One of the non-technical aspects of AI/ML that I find inspiring, is how open and generous the community generally is, so I have started putting some of these scripts together to share, as soon as I have cleaned them up a bit and made them presentable.

Posted on

Game Development Should Not Be a Marathon of Sprints

You probably know the quote, “it’s not a sprint, it’s a marathon.” Usually used in reference to long term thinking, resource management and prudent planning. However, if you work in games a “sprint” will likely remind you of how work is structured into similarly named time blocks of 2-4 weeks1. One sprint follows the other in a never ending line. So, while the saying suggests that you need to pace yourself, measure your efforts towards the end goal, and save your strength for when it matters, reality puts you in a state of constant sprinting.

Of course, at that point it’s not really a sprint at all. At that point, you are a jogger, pretending to sprint to keep up appearances.

The basic idea is rooted in Agile; a much-interpreted and widely adopted ideology within the realm of project management. Especially in game and software development. Many will tell you that Agile was never meant to be constant sprinting, and I agree, however this argument fails to address that it is what we’ve ended up with regardless. As with all ideologies, there are multiple camps. And like all ideologies, there are good intentions and useful ideas to be found, but putting too much weight on the wrong parts leads to needlessly rigid, misguided, and sometimes extreme implementations.

The Cult of Scrum

One such Agile implementation, or framework, is Scrum. It comes with several recurring “rituals” conducted by a “Scrum Master”. Yes, those are the actual terms, and yes, it has a cult vibe. The core idea combines accountability with planning, enforced by regular meetings (aka rituals) where the PM (aka scrum master) walks through the steps outlined in the Scrum guidelines. The regular meetings include: Sprint planning and retros, daily stand up, and backlog grooming. Pretty standard stuff, once you strip the ridiculous terminology away. So, what’s the problem?

The problem is that it doesn’t work as intended. It kind of works with a lot of tweaking, overhead and even fudging. The larger the game studio, the more of this you are likely to see.

One-Size-Fits-Nobody

Different teams work with completely different mind sets, expectations and needs. Squeezing all this diversity into a highly structured approach causes frustration, usually because of the time spent on meetings vs. production, the minutia of updating work items, or a feeling that the PM/management decision making is out of touch. Scrum tries to be one-size-fits-all which isn’t well suited for highly specialized teams. Environment artists approach their work differently than database engineers do theirs and forcing both to follow the same strict process should be an obvious, terrible decision. Yet, here we are, and it’s somehow the industry standard.

Now, if all your work fits nicely into 2-week pockets, you may not be able to relate to this. Small indie studios are also much less affected, generally speaking. If you can’t relate to anything in this article, I say – without any intended sarcasm – good for you! However, I hope you can appreciate that others may have a different take.

In my experience, Scrum unintentionally promotes the worst parts of constant sprinting, mainly due to its high number of meetings.

It’s All in the Numbers

Eventually, teams begin planning work so individual components can be finished within a sprint window, regardless of whether it matches reality. Some prefer writing out work items in the system that matches the work done, after it was finished. That way your time estimates are always accurate. As dumb as this sounds to outsiders, all this can usually be boiled down to making numbers look good on a report.

Once focus drifts to prioritize reporting, the sprints become more about completion rates, burn down charts, and making the team look like they know how to plan their work without taking on too much or coming off as lazy. After all, poor performance might attract unwanted attention when the next round of layoffs come around. Sadly, many team managers and producers out there are stuck with doing this dance as one of their main responsibilities.

What ends up happening, is that each team tweaks the process to fit their needs and get the work done, while trying to look as good as possible on the report. A large portion of work becomes about communicating progress in a way that fits into this pattern and may create a weird competition between teams, where there isn’t a winner per se, but no one wants to be the badly organized team on the bottom, or the team that over-plans and under-delivers. Whether or not that is the case comes down to those who receive said reports, and the culture they are fostering.

If you are reading this, and you’re in this group of stakeholders: studio leadership and upper management, I urge you to give this some thought. How much of this applies at our studio? Is it happening, and I am not even aware of it? What can be done?

Making Sprints Matter

The original idea of the sprint, is that you take a portion of time and spend that time narrowly focusing on a limited scope. You put in all your efforts toward achieving clearly defined goals within this allotted time, often a deliverable of some kind, after which you go back to work as usual. The finish line is not supposed to also start the next sprint, unless it’s a relay and a whole new team is standing by to take over. If that is really the case, I hope your documentation is good. Bottom line is this: No one can keep sprinting and pushing forever and still give 100%. That is how you end up with people trying to game the system, play politics, or simply burning out.

The 6:2 Ratio Split

What if teams worked for 6 weeks in a more freeform structure, where they don’t have to stop and think about story points and sprint retros, and in the meantime, the PM team alongside the team leads, can plan an actual 2 week sprint with a theme that fits where we are in the process. Everyone executes the sprint together, pushing to reach the goal, and when it’s done take a moment to celebrate it, and talk about lessons learned and all that good stuff. The stuff which tends to feel forced and easily forgotten, when sprint retros are held every 2 weeks.

Note that the non-sprinting weeks aren’t down time, either. There is still work to be done, and likely deadlines that fall outside of sprints. The difference is that while not sprinting, the individual teams are free to approach the work in any way that works best for them.

With a 6:2 split, you end up with 6 sprints per year, one every other month, compared to 26 if you never stop sprinting. However, you’re actually sprinting and not simply jogging along. I am not a behavioral scientist but I bet, if you compared this approach with constant sprinting, the productivity would be similar but with higher satisfaction among the majority of the people doing the work. I would even go so far as to say, using Scrum might work too if you don’t keep sprinting.

Isn’t That Crunch With a Different Name?

No. First of all, crunch is where workers are pushed to work long hours, sometimes sleeping in the office, in an effort to meet some insane and unchangeable deadline. Crunch is the result of poor planning at the highest level, usually starting with deadlines being set based on publishing and marketing rollout, without proper consideration for setbacks, individual team capacity, cross-team interdependency, vendor availability, recruiting and all the other variables involved.

The Sprint Purpose

The purpose of the sprint is to push for completion of specifically scoped work. Pushing, in this context, does not refer to pushing your personal life out of the way. Instead, if refers to putting other work-related projects and tasks on hold for this limited time, and instead put those efforts toward the sprint goal. And again, this only works if you are not constantly sprinting, which is why I suggest a split.

Experiment With the Format

Please don’t take the 6:2 ratio as some kind of rule, it’s just the ratio I would start with. If a different split ratio works better for your studio or project, by all means go for it – my point is simply this: Constant sprinting is bad and I encourage you to try a different approach.

One variant that I have seen a few times, which I don’t recommend, is that every so often a sprint is removed and developers are encouraged to work on pet projects or those nice-to-have things they don’t usually have time for. But it rarely works out like that in my experience. Instead, these non-sprints become unstructured catching-up-and-backlog-cleanup sprints. Best case scenario, it will feel like a short moment to re-orient before the next leg of the sprint-marathon begins. In this variant, the ratio is reversed and looking more like 2:10.

I would love to know if you have tried a split ratio or other alternative to constant sprinting, and what your experience was with it.

The Plug

Incidentally, I am working on a book on game production, where I will be exploring this and similar topics. In it, I tie my thoughts to anecdotes and personal experiences for detail and context. There is no link or anything yet, but feel free to ask questions and make suggestions.

  1. Sprint duration varies from studio to studio. For the purpose of this article, we will count a standard sprint as 2 weeks long. ↩︎