I’ve been attempting to grasp one of many newest AI coding buzzword: Spec-driven improvement (SDD). I checked out three of the instruments that label themselves as SDD instruments and tried to untangle what it means, as of now.
Definition
Like with many rising phrases on this fast-paced house, the definition of “spec-driven improvement” (SDD) remains to be in flux. Right here’s what I can collect from how I’ve seen it used to this point: Spec-driven improvement means writing a “spec” earlier than writing code with AI (“documentation first”). The spec turns into the supply of reality for the human and the AI.
GitHub: “On this new world, sustaining software program means evolving specs. […] The lingua franca of improvement strikes to a better degree, and code is the last-mile method.”
Tessl: “A improvement method the place specs — not code — are the first artifact. Specs describe intent in structured, testable language, and brokers generate code to match them.”
After wanting over the usages of the time period, and a few of the instruments that declare to be implementing SDD, it appears to me that in actuality, there are a number of implementation ranges to it:
- Spec-first: A nicely thought-out spec is written first, after which used within the AI-assisted improvement workflow for the duty at hand.
- Spec-anchored: The spec is stored even after the duty is full, to proceed utilizing it for evolution and upkeep of the respective characteristic.
- Spec-as-source: The spec is the principle supply file over time, and solely the spec is edited by the human, the human by no means touches the code.
All SDD approaches and definitions I’ve discovered are spec-first, however not all attempt to be spec-anchored or spec-as-source. And infrequently it’s left obscure or completely open what the spec upkeep technique over time is supposed to be.
What’s a spec?
The important thing query by way of definitions after all is: What’s a spec? There doesn’t appear to be a normal definition, the closest I’ve seen to a constant definition is the comparability of a spec to a “Product Necessities Doc”.
The time period is sort of overloaded in the intervening time, right here is my try at defining what a spec is:
A spec is a structured, behavior-oriented artifact – or a set of associated artifacts – written in pure language that expresses software program performance and serves as steering to AI coding brokers. Every variant of spec-driven improvement defines their method to a spec’s construction, degree of element, and the way these artifacts are organized inside a venture.
There’s a helpful distinction to be made I believe between specs and the extra normal context paperwork for a codebase. That normal context are issues like guidelines information, or excessive degree descriptions of the product and the codebase. Some instruments name this context a reminiscence financial institution, in order that’s what I’ll use right here. These information are related throughout all AI coding classes within the codebase, whereas specs solely related to the duties that really create or change that exact performance.
It seems to be fairly time-consuming to guage SDD instruments and approaches in a method that will get near actual utilization. You would need to attempt them out with completely different sizes of issues, greenfield, brownfield, and actually take the time to assessment and revise the intermediate artifacts with greater than only a cursory look. As a result of as GitHub’s weblog publish about spec-kit says: “Crucially, your position isn’t simply to steer. It’s to confirm. At every part, you mirror and refine.”
For 2 of the three instruments I attempted it additionally appears to be much more work to introduce them into an current codebase, due to this fact making it even tougher to guage their usefulness for brownfield codebases. Till I hear utilization stories from folks utilizing them for a time period on a “actual” codebase, I nonetheless have lots of open questions on how this works in actual life.
That being stated – let’s get into three of those instruments. I’ll share an outline of how they work first (or reasonably how I believe they work), and can preserve my observations and questions for the tip. Be aware that these instruments are very quick evolving, so they may have already modified since I used them in September.
Kiro
Kiro is the only (or most light-weight) one of many three I attempted. It appears to be principally spec-first, all of the examples I’ve discovered use it for a job, or a consumer story, with no point out of the way to use the necessities doc in a spec-anchored method over time, throughout a number of duties.
Workflow: Necessities → Design → Duties
Every workflow step is represented by one markdown doc, and Kiro guides you thru these 3 workflow steps inside its VS Code primarily based distribution.
Necessities: Structured as a listing of necessities, the place every requirement represents a “Consumer Story” (in “As a…” format) with acceptance standards (in “GIVEN… WHEN… THEN…” format)
Design: In my try, the design doc consisted of the sections seen within the screenshot under. I solely have the outcomes of certainly one of my makes an attempt nonetheless, so I’m undecided if this can be a constant construction, or if it adjustments relying on the duty.
Duties: An inventory of duties that hint again to the requirement numbers, and that get some additional UI parts to run duties one after the other, and assessment adjustments per job.
Kiro additionally has the idea of a reminiscence financial institution, they name it “steering”. Its contents are versatile, and their workflow doesn’t appear to depend on any particular information being there (I made my utilization makes an attempt earlier than I even found the steering part). The default topology created by Kiro whenever you ask it to generate steering paperwork is product.md, construction.md, tech.md.
Spec-kit
Spec-kit is GitHub’s model of SDD. It’s distributed as a CLI that may create workspace setups for a variety of frequent coding assistants. As soon as that construction is about up, you work together with spec-kit through slash instructions in your coding assistant. As a result of all of its artifacts are put proper into your workspace, that is essentially the most customizable one of many three instruments mentioned right here.
Workflow: Structure → 𝄆 Specify → Plan → Duties 𝄇
Spec-kit’s reminiscence financial institution idea is a prerequisite for the spec-driven method. They name it a structure. The structure is meant to include the excessive degree rules which can be “immutable” and will at all times be utilized, to each change. It’s mainly a really highly effective guidelines file that’s closely utilized by the workflow.
In every of the workflow steps (specify, plan, duties), spec-kit instantiates a set of information and prompts with the assistance of a bash script and a few templates. The workflow then makes heavy use of checklists inside the information, to trace vital consumer clarifications, structure violations, analysis duties, and many others. They’re like a “definition of completed” for every workflow step (although interpreted by AI, so there is no such thing as a 100% assure that they are going to be revered).
Beneath is an outline as an example the file topology I noticed in spec-kit. Be aware how one spec is made up of many information.
At first look, GitHub appears to be aspiring to a spec-anchored method (“That’s why we’re rethinking specs — not as static paperwork, however as residing, executable artifacts that evolve with the venture. Specs develop into the shared supply of reality. When one thing doesn’t make sense, you return to the spec; when a venture grows complicated, you refine it; when duties really feel too massive, you break them down.”) Nevertheless, spec-kit creates a department for each spec that will get created, which appears to point that they see a spec as a residing artifact for the lifetime of a change request, not the lifetime of a characteristic. This neighborhood dialogue is speaking about this confusion. It makes me assume that spec-kit remains to be what I might name spec-first solely, not spec-anchored over time.
Tessl Framework
(Nonetheless in non-public beta)
Like spec-kit, the Tessl Framework is distributed as a CLI that may create all of the workspace and config construction for quite a lot of coding assistants. The CLI command additionally doubles as an MCP server.
Tessl is the one certainly one of these three instruments that explicitly aspires to a spec-anchored method, and is even exploring the spec-as-source degree of SDD. A Tessl spec can function the principle artifact that’s being maintained and edited, with the code even marked with a remark on the high saying // GENERATED FROM SPEC - DO NOT EDIT. That is presently a 1:1 mapping between spec and code information, i.e. one spec interprets into one file within the codebase. However Tessl remains to be in beta and they’re experimenting with completely different variations of this, so I can think about that this method is also taken on a degree the place one spec maps to a code element with a number of information. It stays to be seen what the alpha product will assist. (The Tessl workforce themselves see their framework as one thing that’s extra sooner or later than their present public product, the Tessl Registry.)
Right here is an instance of a spec that I had the Tessl CLI reverse engineer (tessl doc --code ...js) from a JavaScript file in an current codebase:
Tags like @generate or @check appear to inform Tessl what to generate. The API part reveals the thought of defining a minimum of the interfaces that get uncovered to different components of the codebase within the spec, presumably to guarantee that these extra essential components of the generated element are totally beneath the management of the maintainer. Operating tessl construct for this spec generates the corresponding JavaScript code file.
Placing the specs for spec-as-source at a fairly low abstraction degree, per code file, most likely reduces quantity of steps and interpretations the LLM has to do, and due to this fact the prospect of errors. Even at this low abstraction degree I’ve seen the non-determinism in motion although, once I generated code a number of instances from the identical spec. It was an attention-grabbing train to iterate on the spec and make it increasingly particular to extend the repeatability of the code era. That course of jogged my memory of a few of the pitfalls and challenges of writing an unambiguous and full specification.
Observations and questions
These three instruments are all labelling themselves as implementations of spec-driven improvement, however they’re fairly completely different from one another. In order that’s the very first thing to bear in mind when speaking about SDD, it’s not only one factor.
One workflow to suit all sizes?
Kiro and spec-kit present one opinionated workflow every, however I’m fairly certain that neither of them is appropriate for almost all of actual life coding issues. Specifically, it’s not fairly clear to me how they might cater to sufficient completely different downside sizes to be typically relevant.
After I requested Kiro to repair a small bug (it was the identical one I used prior to now to attempt Codex), it shortly turned clear that the workflow was like utilizing a sledgehammer to crack a nut. The necessities doc turned this small bug into 4 “consumer tales” with a complete of 16 acceptance standards, together with gems like “Consumer story: As a developer, I need the transformation operate to deal with edge instances gracefully, in order that the system stays strong when new class codecs are launched.”
I had an identical problem once I used spec-kit, I wasn’t fairly certain what measurement of downside to make use of it for. Obtainable tutorials are often primarily based on creating an software from scratch, as a result of that’s best for a tutorial. One of many use instances I ended up attempting was a characteristic that may be a 3-5 level story on certainly one of my previous groups. The characteristic trusted lots of code that was already there, it was supposed to construct an outline modal that summarised a bunch of information from an current dashboard. With the quantity of steps spec-kit took, and the quantity of markdown information it created for me to assessment, this once more felt like overkill for the dimensions of the issue. It was a much bigger downside than the one I used with Kiro, but additionally a way more elaborate workflow. I by no means even completed the total implementation, however I believe in the identical time it took me to run and assessment the spec-kit outcomes I may have carried out the characteristic with “plain” AI-assisted coding, and I might have felt far more in management.
An efficient SDD instrument would on the very least have to supply flexibility for just a few completely different core workflows, for various sizes and kinds of adjustments.
Reviewing markdown over reviewing code?
As simply talked about, and as you’ll be able to see within the description of the instrument above, spec-kit created a LOT of markdown information for me to assessment. They had been repetitive, each with one another, and with the code that already existed. Some contained code already. Total they had been simply very verbose and tedious to assessment. In Kiro it was just a little simpler, as you solely get 3 information, and it’s extra intuitive to grasp the psychological mannequin of “necessities > design > duties”. Nevertheless, as talked about, Kiro additionally was method too verbose for the small bug I used to be asking it to repair.
To be trustworthy, I’d reasonably assessment code than all these markdown information. An efficient SDD instrument must present an excellent spec assessment expertise.
False sense of management?
Even with all of those information and templates and prompts and workflows and checklists, I incessantly noticed the agent in the end not comply with all of the directions. Sure, the context home windows are actually bigger, which is commonly talked about as one of many enablers of spec-driven improvement. However simply because the home windows are bigger, doesn’t imply that AI will correctly decide up on every little thing that’s in there.
For instance: Spec-kit has a analysis step someplace throughout planning, and it did lots of analysis on the present code and what’s already there, which was nice as a result of I requested it so as to add a characteristic that constructed on high of current code. However in the end the agent ignored the notes that these had been descriptions of current lessons, it simply took them as a brand new specification and generated them over again, creating duplicates. However I didn’t solely see examples of ignoring directions, I additionally noticed the agent go method overboard as a result of it was too eagerly following directions (e.g. one of many structure articles).
The previous has proven that the easiest way for us to remain accountable for what we’re constructing are small, iterative steps, so I’m very skeptical that a lot of up-front spec design is a good suggestion, particularly when it’s overly verbose. An efficient SDD instrument must cater to an iterative method, however small work packages virtually appear counter to the thought of SDD.
How one can successfully separate useful from technical spec?
It’s a frequent thought in SDD to be intentional concerning the separation between useful spec and technical implementation. The underlying aspiration I assume is that in the end, we may have AI fill in all of the solutioning and particulars, and swap to completely different tech stacks with the identical spec.
In actuality, once I was attempting spec-kit, I incessantly received confused when to remain on the useful degree, and when it was time so as to add technical particulars. The tutorial and documentation additionally weren’t fairly per it, there appear to be completely different interpretations of what “purely useful” actually means. And once I assume again on the numerous, many consumer tales I’ve learn in my profession that weren’t correctly separating necessities from implementation, I don’t assume we’ve monitor document as a career to do that nicely.
Who’s the goal consumer?
Most of the demos and tutorials for spec-driven improvement instruments embody issues like defining product and have objectives, they even incorporate phrases like “consumer story”. The concept right here could be to make use of AI as an enabler for cross-skilling, and have builders take part extra closely in necessities evaluation? Or have builders pair with product folks after they work on this workflow? None of that is made specific although, it’s introduced as a given {that a} developer would do all this evaluation.
By which case I might ask myself once more, what downside measurement and kind is SDD meant for? Most likely not for giant options which can be nonetheless very unclear, as absolutely that may require extra specialist product and necessities expertise, and plenty of different steps like analysis and stakeholder involvement?
Spec-anchored and spec-as-source: Are we studying from the previous?
Whereas many individuals draw analogies between SDD and TDD or BDD, I believe one other necessary parallel to take a look at for spec-as-source particularly is MDD (model-driven improvement). I labored on just a few initiatives at first of my profession that closely used MDD, and I stored being reminded about that once I was attempting out the Tessl Framework. The fashions in MDD had been mainly the specs, albeit not in pure language, however expressed in e.g. customized UML or a textual DSL. We constructed customized code mills to show these specs into code.
In the end, MDD by no means took off for enterprise purposes, it sits at an ungainly abstraction degree and simply creates an excessive amount of overhead and constraints. However LLMs take a few of the overhead and constraints of MDD away, so there’s a new hope that we will now lastly concentrate on writing specs and simply generate code from them. With LLMs, we aren’t constrained by a predefined and parseable spec language anymore, and we don’t should construct elaborate code mills. The worth for that’s LLMs’ non-determinism after all. And the parseable construction additionally had upsides that we’re dropping now: We may present the spec writer with lots of instrument assist to write down legitimate, full and constant specs. I ponder if spec-as-source, and even spec-anchoring, would possibly find yourself with the downsides of each MDD and LLMs: Inflexibility and non-determinism.
To be clear, I’m not nostalgic about my MDD expertise prior to now and saying “we’d as nicely carry that again”. However we should always look to code-from-spec makes an attempt prior to now to study from them once we discover spec-driven at present.
Conclusions
In my private utilization of AI-assisted coding, I additionally typically spend time on fastidiously crafting some type of spec first to offer to the coding agent. So the final precept of spec-first is certainly priceless in lots of conditions, and the completely different approaches of the way to construction that spec are very wanted. They’re among the many high most incessantly requested questions I hear in the intervening time from practitioners: “How do I construction my reminiscence financial institution?”, “How do I write specification and design doc for AI?”.
However the time period “spec-driven improvement” isn’t very nicely outlined but, and it’s already semantically subtle. I’ve even lately heard folks use “spec” mainly as a synonym for “detailed immediate”.
Concerning the instruments I’ve tried, I’ve listed a lot of my questions on their actual world usefulness right here. I ponder if a few of them try to feed AI brokers with our current workflows too actually, in the end amplifying current challenges like assessment overload and hallucinations. Particularly with the extra elaborate approaches that create a lot of information, I can’t assist however consider the German compound phrase “Verschlimmbesserung”: Are we making one thing worse within the try of constructing it higher?
