Learn how to Write a Good Spec for AI Brokers – O’Reilly

This publish first appeared on Addy Osmani’s Elevate Substack publication and is being republished right here with the creator’s permission.

TL;DR: Intention for a transparent spec protecting simply sufficient nuance (this will embody construction, fashion, testing, boundaries. . .) to information the AI with out overwhelming it. Break giant duties into smaller ones versus maintaining the whole lot in a single giant immediate. Plan first in read-only mode, then execute and iterate constantly.

“I’ve heard rather a lot about writing good specs for AI brokers, however haven’t discovered a strong framework but. I might write a spec that rivals an RFC, however sooner or later the context is simply too giant and the mannequin breaks down.”

Many builders share this frustration. Merely throwing a large spec at an AI agent doesn’t work—context window limits and the mannequin’s “consideration funds” get in the way in which. The secret is to put in writing sensible specs: paperwork that information the agent clearly, keep inside sensible context sizes, and evolve with the mission. This information distills finest practices from my use of coding brokers together with Claude Code and Gemini CLI right into a framework for spec-writing that retains your AI brokers targeted and productive.

We’ll cowl 5 ideas for excellent AI agent specs, every beginning with a bolded takeaway.

1. Begin with a Excessive-Stage Imaginative and prescient and Let the AI Draft the Particulars

Kick off your mission with a concise high-level spec, then have the AI broaden it into an in depth plan.

As a substitute of overengineering upfront, start with a transparent purpose assertion and some core necessities. Deal with this as a “product transient” and let the agent generate a extra elaborate spec from it. This leverages the AI’s power in elaboration when you preserve management of the course. This works properly except you already really feel you could have very particular technical necessities that have to be met from the beginning.

Why this works: LLM-based brokers excel at fleshing out particulars when given a strong high-level directive, however they want a transparent mission to keep away from drifting astray. By offering a brief define or goal description and asking the AI to supply a full specification (e.g., a spec.md), you create a persistent reference for the agent. Planning prematurely issues much more with an agent: You possibly can iterate on the plan first, then hand it off to the agent to put in writing the code. The spec turns into the primary artifact you and the AI construct collectively.

Sensible method: Begin a brand new coding session by prompting 

You're an AI software program engineer. Draft an in depth specification for
[project X] protecting targets, options, constraints, and a step-by-step plan.

Preserve your preliminary immediate high-level: e.g., “Construct an internet app the place customers can
monitor duties (to-do record), with person accounts, a database, and a easy UI.”

The agent may reply with a structured draft spec: an outline, function record, tech stack options, information mannequin, and so forth. This spec then turns into the “supply of fact” that each you and the agent can refer again to. GitHub’s AI workforce promotes spec-driven growth the place “specs turn out to be the shared supply of fact…residing, executable artifacts that evolve with the mission.” Earlier than writing any code, evaluation and refine the AI’s spec. Ensure that it aligns together with your imaginative and prescient and proper any hallucinations or off-target particulars.

Use Plan Mode to implement planning-first: Instruments like Claude Code supply a Plan Mode that restricts the agent to read-only operations—it could analyze your codebase and create detailed plans however gained’t write any code till you’re prepared. That is excellent for the planning part: Begin in Plan Mode (Shift+Tab in Claude Code), describe what you wish to construct, and let the agent draft a spec whereas exploring your present code. Ask it to make clear ambiguities by questioning you in regards to the plan. Have it evaluation the plan for structure, finest practices, safety dangers, and testing technique. The purpose is to refine the plan till there’s no room for misinterpretation. Solely then do you exit Plan Mode and let the agent execute. This workflow prevents the widespread lure of leaping straight into code era earlier than the spec is strong.

Use the spec as context: As soon as accepted, save this spec (e.g., as SPEC.md) and feed related sections into the agent as wanted. Many builders utilizing a powerful mannequin do precisely this. The spec file persists between classes, anchoring the AI each time work resumes on the mission. This mitigates the forgetfulness that may occur when the dialog historical past will get too lengthy or when it’s important to restart an agent. It’s akin to how one would use a product necessities doc (PRD) in a workforce: a reference that everybody (human or AI) can seek the advice of to remain on monitor. Skilled people usually “write good documentation first and the mannequin could possibly construct the matching implementation from that enter alone” as one engineer noticed. The spec is that documentation.

Preserve it purpose oriented: A high-level spec for an AI agent ought to give attention to what and why greater than the nitty-gritty how (at the very least initially). Consider it just like the person story and acceptance standards: Who’s the person? What do they want? What does success appear to be? (For instance, “Consumer can add, edit, full duties; information is saved persistently; the app is responsive and safe.”) This retains the AI’s detailed spec grounded in person wants and consequence, not simply technical to-dos. Because the GitHub Spec Package docs put it, present a high-level description of what you’re constructing and why, and let the coding agent generate an in depth specification specializing in person expertise and success standards. Beginning with this big-picture imaginative and prescient prevents the agent from dropping sight of the forest for the bushes when it later will get into coding.

2. Construction the Spec Like a Skilled PRD (or SRS)

Deal with your AI spec as a structured doc (PRD) with clear sections, not a unfastened pile of notes.

Many builders deal with specs for brokers very like conventional product requirement paperwork (PRDs) or system design docs: complete, well-organized, and straightforward for a “literal-minded” AI to parse. This formal method provides the agent a blueprint to comply with and reduces ambiguity.

The six core areas

GitHub’s evaluation of over 2,500 agent configuration information revealed a transparent sample: The simplest specs cowl six areas. Use this as a guidelines for completeness:

1. Instructions: Put executable instructions early—not simply device names however full instructions with flags: npm check, pytest -v, npm run construct. The agent will reference these always.
2. Testing: Learn how to run assessments, what framework you employ, the place check information stay, and what protection expectations exist.
3. Venture construction: The place supply code lives, the place assessments go, the place docs belong. Be express: “src/ for utility code, assessments/ for unit assessments, docs/ for documentation.”
4. Code fashion: One actual code snippet exhibiting your fashion beats three paragraphs describing it. Embody naming conventions, formatting guidelines, and examples of excellent output.
5. Git workflow: Department naming, commit message format, PR necessities. The agent can comply with these in the event you spell them out.
6. Boundaries: What the agent ought to by no means contact—secrets and techniques, vendor directories, manufacturing configs, particular folders. “By no means commit secrets and techniques” was the only most typical useful constraint within the GitHub examine.

Be particular about your stack: Say “React 18 with TypeScript, Vite, and Tailwind CSS,” not “React mission.” Embody variations and key dependencies. Imprecise specs produce obscure code.

Use a constant format: Readability is king. Many devs use Markdown headings and even XML-like tags within the spec to delineate sections as a result of AI fashions deal with well-structured textual content higher than free-form prose. For instance, you may construction the spec as:

# Venture Spec: My workforce's duties app


## Goal
- Construct an internet app for small groups to handle duties...


## Tech Stack
- React 18+, TypeScript, Vite, Tailwind CSS
- Node.js/Categorical backend, PostgreSQL, Prisma ORM


## Instructions
- Construct: `npm run construct` (compiles TypeScript, outputs to dist/)
- Check: `npm check` (runs Jest, should cross earlier than commits)
- Lint: `npm run lint --fix` (auto-fixes ESLint errors)


## Venture Construction
- `src/` – Software supply code
- `assessments/` – Unit and integration assessments
- `docs/` – Documentation


## Boundaries
- ✅ At all times: Run assessments earlier than commits, comply with naming conventions
- ⚠️ Ask first: Database schema modifications, including dependencies
- 🚫 By no means: Commit secrets and techniques, edit node_modules/, modify CI config

This degree of group not solely helps you assume clearly but additionally helps the AI discover data. Anthropic engineers suggest organizing prompts into distinct sections (like , , , and so forth.) for precisely this purpose: It provides the mannequin robust cues about which data is which. And keep in mind, “minimal doesn’t essentially imply quick”—don’t shrink back from element within the spec if it issues, however maintain it targeted.

Combine specs into your toolchain: Deal with specs as “executable artifacts” tied to model management and CI/CD. The GitHub Spec Package makes use of a four-phase gated workflow that makes your specification the middle of your engineering course of. As a substitute of writing a spec and setting it apart, the spec drives the implementation, checklists, and job breakdowns. Your main function is to steer; the coding agent does the majority of the writing. Every part has a selected job, and also you don’t transfer to the following one till the present job is absolutely validated:

1. Specify: You present a high-level description of what you’re constructing and why, and the coding agent generates an in depth specification. This isn’t about technical stacks or app design—it’s about person journeys, experiences, and what success seems to be like. Who will use this? What downside does it resolve? How will they work together with it? Consider it as mapping the person expertise you wish to create, and letting the coding agent flesh out the small print. This turns into a residing artifact that evolves as you study extra.

2. Plan: Now you get technical. You present your required stack, structure, and constraints, and the coding agent generates a complete technical plan. If your organization standardizes on sure applied sciences, that is the place you say so. When you’re integrating with legacy methods or have compliance necessities, all of that goes right here. You possibly can ask for a number of plan variations to check approaches. When you make inner docs obtainable, the agent can combine your architectural patterns immediately into the plan.

3. Duties: The coding agent takes the spec and plan and breaks them into precise work—small, reviewable chunks that every resolve a selected piece of the puzzle. Every job must be one thing you possibly can implement and check in isolation, virtually like test-driven growth to your AI agent. As a substitute of “construct authentication,” you get concrete duties like “create a person registration endpoint that validates electronic mail format.”

4. Implement: Your coding agent tackles duties one after the other (or in parallel). As a substitute of reviewing thousand-line code dumps, you evaluation targeted modifications that resolve particular issues. The agent is aware of what to construct (specification), how one can construct it (plan), and what to work on (job). Crucially, your function is to confirm at every part: Does the spec seize what you need? Does the plan account for constraints? Are there edge instances the AI missed? The method builds in checkpoints so that you can critique, spot gaps, and course-correct earlier than shifting ahead.

This gated workflow prevents what Willison calls “home of playing cards code”: fragile AI outputs that collapse below scrutiny. Anthropic’s Abilities system presents an identical sample, letting you outline reusable Markdown-based behaviors that brokers invoke. By embedding your spec in these workflows, you make sure the agent can’t proceed till the spec is validated, and modifications propagate robotically to job breakdowns and assessments.

Take into account brokers.md for specialised personas: For instruments like GitHub Copilot, you possibly can create brokers.md information that outline specialised agent personas—a @docs-agent for technical writing, a @test-agent for QA, a @security-agent for code evaluation. Every file acts as a targeted spec for that persona’s habits, instructions, and bounds. That is significantly helpful if you need totally different brokers for various duties reasonably than one general-purpose assistant.

Design for agent expertise (AX): Simply as we design APIs for developer expertise (DX), take into account designing specs for “agent expertise.” This implies clear, parseable codecs: OpenAPI schemas for any APIs the agent will devour, llms.txt information that summarize documentation for LLM consumption, and express sort definitions. The Agentic AI Basis (AAIF) is standardizing protocols like MCP (Mannequin Context Protocol) for device integration. Specs that comply with these patterns are simpler for brokers to devour and act on reliably.

PRD versus SRS mindset: It helps to borrow from established documentation practices. For AI agent specs, you’ll usually mix these into one doc (as illustrated above), however protecting each angles serves you properly. Writing it like a PRD ensures you embody user-centric context (“the why behind every function”) so the AI doesn’t optimize for the flawed factor. Increasing it like an SRS ensures you nail down the specifics the AI might want to really generate right code (like what database or API to make use of). Builders have discovered that this additional upfront effort pays off by drastically decreasing miscommunications with the agent later.

Make the spec a “residing doc”: Don’t write it and overlook it. Replace the spec as you and the agent make choices or uncover new data. If the AI needed to change the information mannequin otherwise you determined to chop a function, replicate that within the spec so it stays the bottom fact. Consider it as version-controlled documentation. In spec-driven workflows, the spec drives implementation, assessments, and job breakdowns, and also you don’t transfer to coding till the spec is validated. This behavior retains the mission coherent, particularly in the event you or the agent step away and are available again later. Bear in mind, the spec isn’t only for the AI—it helps you because the developer preserve oversight and make sure the AI’s work meets the actual necessities.

3. Break Duties into Modular Prompts and Context, Not One Massive Immediate

Divide and conquer: Give the AI one targeted job at a time reasonably than a monolithic immediate with the whole lot directly.

Skilled AI engineers have realized that attempting to stuff your complete mission (all necessities, all code, all directions) right into a single immediate or agent message is a recipe for confusion. Not solely do you danger hitting token limits; you additionally danger the mannequin dropping focus as a result of “curse of directions”—too many directives inflicting it to comply with none of them properly. The answer is to design your spec and workflow in a modular means, tackling one piece at a time and pulling in solely the context wanted for that piece.

The curse of an excessive amount of context/directions: Analysis has confirmed what many devs anecdotally noticed: as you pile on extra directions or information into the immediate, the mannequin’s efficiency in adhering to every one drops considerably. One examine dubbed this the “curse of directions”, exhibiting that even GPT-4 and Claude battle when requested to fulfill many necessities concurrently. In sensible phrases, in the event you current 10 bullet factors of detailed guidelines, the AI may obey the primary few and begin overlooking others. The higher technique is iterative focus. Tips from trade counsel decomposing complicated necessities into sequential, easy directions as a finest follow. Focus the AI on one subproblem at a time, get that carried out, then transfer on. This retains the standard excessive and errors manageable.

Divide the spec into phases or parts: In case your spec doc could be very lengthy or covers numerous floor, take into account splitting it into elements (both bodily separate information or clearly separate sections). For instance, you may need a piece for “backend API spec” and one other for “frontend UI spec.” You don’t must at all times feed the frontend spec to the AI when it’s engaged on the backend, and vice versa. Many devs utilizing multi-agent setups even create separate brokers or subprocesses for every half (e.g., one agent works on database/schema, one other on API logic, one other on frontend—every with the related slice of the spec). Even in the event you use a single agent, you possibly can emulate this by copying solely the related spec part into the immediate for that job. Keep away from context overload: Don’t combine authentication duties with database schema modifications in a single go, because the DigitalOcean AI information warns. Preserve every immediate tightly scoped to the present purpose.

Prolonged TOC/summaries for big specs: One intelligent approach is to have the agent construct an prolonged desk of contents with summaries for the spec. That is primarily a “spec abstract” that condenses every part into a couple of key factors or key phrases, and references the place particulars may be discovered. For instance, in case your full spec has a piece on safety necessities spanning 500 phrases, you may need the agent summarize it to: “Safety: Use HTTPS, defend API keys, implement enter validation (see full spec §4.2).” By making a hierarchical abstract within the planning part, you get a fowl’s-eye view that may keep within the immediate, whereas the positive particulars stay offloaded except wanted. This prolonged TOC acts as an index: The agent can seek the advice of it and say, “Aha, there’s a safety part I ought to take a look at,” and you’ll then present that part on demand. It’s just like how a human developer skims a top level view after which flips to the related web page of a spec doc when engaged on a selected half.

To implement this, you possibly can immediate the agent after writing the spec: “Summarize the spec above into a really concise define with every part’s key factors and a reference tag.” The outcome is perhaps a listing of sections with one or two sentence summaries. That abstract may be stored within the system or assistant message to information the agent’s focus with out consuming up too many tokens. This hierarchical summarization method is thought to assist LLMs preserve long-term context by specializing in the high-level construction. The agent carries a “psychological map” of the spec.

Make the most of subagents or “expertise” for various spec elements: One other superior method is utilizing a number of specialised brokers (what Anthropic calls subagents or what you may name “expertise”). Every subagent is configured for a selected space of experience and given the portion of the spec related to that space. As an example, you may need a database designer subagent that solely is aware of in regards to the information mannequin part of the spec, and an API coder subagent that is aware of the API endpoints spec. The primary agent (or an orchestrator) can route duties to the suitable subagent robotically.

The profit is every agent has a smaller context window to cope with and a extra targeted function, which might increase accuracy and permit parallel work on unbiased duties. Anthropic’s Claude Code helps this by letting you outline subagents with their very own system prompts and instruments. “Every subagent has a selected function and experience space, makes use of its personal context window separate from the primary dialog, and has a customized system immediate guiding its habits,” as their docs describe. When a job comes up that matches a subagent’s area, Claude can delegate that job to it, with the subagent returning outcomes independently.

Parallel brokers for throughput: Working a number of brokers concurrently is rising as “the following large factor” for developer productiveness. Fairly than ready for one agent to complete earlier than beginning one other job, you possibly can spin up parallel brokers for non-overlapping work. Willison describes this as “embracing parallel coding brokers” and notes it’s “surprisingly efficient, if mentally exhausting.” The secret is scoping duties so brokers don’t step on one another: One agent codes a function whereas one other writes assessments, or separate parts get constructed concurrently. Orchestration frameworks like LangGraph or OpenAI Swarm may help coordinate these brokers, and shared reminiscence by way of vector databases (like Chroma) lets them entry widespread context with out redundant prompting.

Single versus multi-agent: When to make use of every

In follow, utilizing subagents or skill-specific prompts may appear to be: You preserve a number of spec information (or immediate templates)—e.g., SPEC_backend.md, SPEC_frontend.md—and also you inform the AI, “For backend duties, seek advice from SPEC_backend; for frontend duties seek advice from SPEC_frontend.” Or in a device like Cursor/Claude, you really spin up a subagent for every. That is actually extra complicated to arrange than a single-agent loop, but it surely mimics what human builders do: We mentally compartmentalize a big spec into related chunks. (You don’t maintain the entire 50-page spec in your head directly; you recall the half you want for the duty at hand, and have a common sense of the general structure.) The problem, as famous, is managing interdependencies: The subagents should nonetheless coordinate. (The frontend must know the API contract from the backend spec, and so forth.) A central overview (or an “architect” agent) may help by referencing the subspecs and guaranteeing consistency.

Focus every immediate on one job/part: Even with out fancy multi-agent setups, you possibly can manually implement modularity. For instance, after the spec is written, your subsequent transfer is perhaps: “Step 1: Implement the database schema.” You feed the agent the database part of the spec solely, plus any world constraints from the spec (like tech stack). The agent works on that. Then for Step 2, “Now implement the authentication function”, you present the auth part of the spec and possibly the related elements of the schema if wanted. By refreshing the context for every main job, you make sure the mannequin isn’t carrying numerous stale or irrelevant data that might distract it. As one information suggests: “Begin contemporary: start new classes to clear context when switching between main options.” You possibly can at all times remind the agent of crucial world guidelines (from the spec’s constraints part) every time, however don’t shove your complete spec in if it’s not all wanted.

Use in-line directives and code TODOs: One other modularity trick is to make use of your code or spec as an lively a part of the dialog. As an example, scaffold your code with // TODO feedback that describe what must be carried out, and have the agent fill them one after the other. Every TODO primarily acts as a mini-spec for a small job. This retains the AI laser targeted (“implement this particular operate in response to this spec snippet”), and you’ll iterate in a decent loop. It’s just like giving the AI a guidelines merchandise to finish reasonably than the entire guidelines directly.

The underside line: Small, targeted context beats one large immediate. This improves high quality and retains the AI from getting “overwhelmed” by an excessive amount of directly. As one set of finest practices sums up, present “One Process Focus” and “Related data solely” to the mannequin, and keep away from dumping the whole lot all over the place. By structuring the work into modules—and utilizing methods like spec summaries or subspec brokers—you’ll navigate round context dimension limits and the AI’s short-term reminiscence cap. Bear in mind, a well-fed AI is sort of a well-fed operate: Give it solely the inputs it wants for the job at hand.

4. Construct in Self-Checks, Constraints, and Human Experience

Make your spec not only a to-do record for the agent but additionally a information for high quality management—and don’t be afraid to inject your personal experience.

A great spec for an AI agent anticipates the place the AI may go flawed and units up guardrails. It additionally takes benefit of what you understand (area information, edge instances, “gotchas”) so the AI doesn’t function in a vacuum. Consider the spec as each coach and referee for the AI: It ought to encourage the correct method and name out fouls.

Use three-tier boundaries: GitHub’s evaluation of two,500+ agent information discovered that the simplest specs use a three-tier boundary system reasonably than a easy record of don’ts. This provides the agent clearer steerage on when to proceed, when to pause, and when to cease:

✅ At all times do: Actions the agent ought to take with out asking. “At all times run assessments earlier than commits.” “At all times comply with the naming conventions within the fashion information.” “At all times log errors to the monitoring service.”

⚠️ Ask first: Actions that require human approval. “Ask earlier than modifying database schemas.” “Ask earlier than including new dependencies.” “Ask earlier than altering CI/CD configuration.” This tier catches high-impact modifications that is perhaps positive however warrant a human test.

🚫 By no means do: Onerous stops. “By no means commit secrets and techniques or API keys.” “By no means edit node_modules/ or vendor/.” “By no means take away a failing check with out express approval.” “By no means commit secrets and techniques” was the only most typical useful constraint within the examine.

This three-tier method is extra nuanced than a flat record of guidelines. It acknowledges that some actions are at all times secure, some want oversight, and a few are categorically off-limits. The agent can proceed confidently on “At all times” objects, flag “Ask first” objects for evaluation, and hard-stop on “By no means” objects.

Encourage self-verification: One highly effective sample is to have the agent confirm its work in opposition to the spec robotically. In case your tooling permits, you possibly can combine checks like unit assessments or linting that the AI can run after producing code. However even on the spec/immediate degree, you possibly can instruct the AI to double-check (e.g., “After implementing, evaluate the outcome with the spec and ensure all necessities are met. Checklist any spec objects that aren’t addressed.”). This pushes the LLM to replicate on its output relative to the spec, catching omissions. It’s a type of self-audit constructed into the method.

As an example, you may append to a immediate: “(After writing the operate, evaluation the above necessities record and guarantee every is happy, marking any lacking ones).” The mannequin will then (ideally) output the code adopted by a brief guidelines indicating if it met every requirement. This reduces the possibility it forgets one thing earlier than you even run assessments. It’s not foolproof, but it surely helps.

LLM-as-a-Decide for subjective checks: For standards which are arduous to check robotically—code fashion, readability, adherence to architectural patterns—think about using “LLM-as-a-Decide.” This implies having a second agent (or a separate immediate) evaluation the primary agent’s output in opposition to your spec’s high quality pointers. Anthropic and others have discovered this efficient for subjective analysis. You may immediate “Assessment this code for adherence to our fashion information. Flag any violations.” The decide agent returns suggestions that both will get integrated or triggers a revision. This provides a layer of semantic analysis past syntax checks.

Conformance testing: Willison advocates constructing conformance suites—language-independent assessments (usually YAML primarily based) that any implementation should cross. These act as a contract: When you’re constructing an API, the conformance suite specifies anticipated inputs/outputs, and the agent’s code should fulfill all instances. That is extra rigorous than advert hoc unit assessments as a result of it’s derived immediately from the spec and may be reused throughout implementations. Embody conformance standards in your spec’s success part (e.g., “Should cross all instances in conformance/api-tests.yaml”).

Leverage testing within the spec: If attainable, incorporate a check plan and even precise assessments in your spec and immediate stream. In conventional growth, we use TDD or write check instances to make clear necessities—you are able to do the identical with AI. For instance, within the spec’s success standards, you may say, “These pattern inputs ought to produce these outputs…” or “The next unit assessments ought to cross.” The agent may be prompted to run by means of these instances in its head or really execute them if it has that functionality. Willison famous that having a strong check suite is like giving the brokers superpowers: They will validate and iterate rapidly when assessments fail. In an AI coding context, writing a little bit of pseudocode for assessments or anticipated outcomes within the spec can information the agent’s implementation. Moreover, you should utilize a devoted “check agent” in a subagent setup that takes the spec’s standards and constantly verifies the “code agent’s” output.

Deliver your area information: Your spec ought to replicate insights that solely an skilled developer or somebody with context would know. For instance, in the event you’re constructing an ecommerce agent and you understand that “merchandise” and “classes” have a many-to-many relationship, state that clearly. (Don’t assume the AI will infer it—it may not.) If a sure library is notoriously difficult, point out pitfalls to keep away from. Basically, pour your mentorship into the spec. The spec can include recommendation like “If utilizing library X, be careful for reminiscence leak situation in model Y (apply workaround Z).” This degree of element is what turns a median AI output into a very strong answer, since you’ve steered the AI away from widespread traps.

Additionally, when you’ve got preferences or fashion pointers (say, “use useful parts over class parts in React”), encode that within the spec. The AI will then emulate your fashion. Many engineers even embody small examples within the spec (as an illustration, “All API responses must be JSON, e.g., {“error”: “message”} for errors.”). By giving a fast instance, you anchor the AI to the precise format you need.

Minimalism for easy duties: Whereas we advocate thorough specs, a part of experience is figuring out when to maintain it easy. For comparatively easy, remoted duties, an overbearing spec can really confuse greater than assist. When you’re asking the agent to do one thing simple (like “middle a div on the web page”), you may simply say, “Ensure that to maintain the answer concise and don’t add extraneous markup or types.” No want for a full PRD there. Conversely, for complicated duties (like “implement an OAuth stream with token refresh and error dealing with”), that’s if you escape the detailed spec. A great rule of thumb: Modify spec element to job complexity. Don’t underspec a tough downside (the agent will flail or go off-track), however don’t overspec a trivial one (the agent may get tangled or dissipate context on pointless directions).

Keep the AI’s “persona” if wanted: Typically, a part of your spec is defining how the agent ought to behave or reply, particularly if the agent interacts with customers. For instance, if constructing a buyer assist agent, your spec may embody pointers like “Use a pleasant {and professional} tone” and “When you don’t know the reply, ask for clarification or supply to comply with up reasonably than guessing.” These sorts of guidelines (usually included in system prompts) assist maintain the AI’s outputs aligned with expectations. They’re primarily spec objects for AI habits. Preserve them constant and remind the mannequin of them if wanted in lengthy classes. (LLMs can “drift” in fashion over time if not stored on a leash.)

You stay the exec within the loop: The spec empowers the agent, however you stay the last word high quality filter. If the agent produces one thing that technically meets the spec however doesn’t really feel proper, belief your judgement. Both refine the spec or immediately regulate the output. The beauty of AI brokers is that they don’t get offended—in the event that they ship a design that’s off, you possibly can say, “Truly, that’s not what I supposed, let’s make clear the spec and redo it.” The spec is a residing artifact in collaboration with the AI, not a one-time contract you possibly can’t change.

Simon Willison humorously likened working with AI brokers to “a really bizarre type of administration” and even “getting good outcomes out of a coding agent feels uncomfortably near managing a human intern.” It is advisable present clear directions (the spec), guarantee they’ve the mandatory context (the spec and related information), and provides actionable suggestions. The spec units the stage, however monitoring and suggestions throughout execution are key. If an AI was a “bizarre digital intern who will completely cheat in the event you give them an opportunity,” the spec and constraints you write are the way you forestall that dishonest and maintain them on job.

Right here’s the payoff: A great spec doesn’t simply inform the AI what to construct; it additionally helps it self-correct and keep inside secure boundaries. By baking in verification steps, constraints, and your hard-earned information, you drastically enhance the percentages that the agent’s output is right on the primary strive (or at the very least a lot nearer to right). This reduces iterations and people “Why on Earth did it do this?” moments.

5. Check, Iterate, and Evolve the Spec (and Use the Proper Instruments)

Consider spec writing and agent constructing as an iterative loop: check early, collect suggestions, refine the spec, and leverage instruments to automate checks.

The preliminary spec shouldn’t be the top—it’s the start of a cycle. The perfect outcomes come if you frequently confirm the agent’s work in opposition to the spec and regulate accordingly. Additionally, fashionable AI devs use varied instruments to assist this course of (from CI pipelines to context administration utilities).

Steady testing: Don’t wait till the top to see if the agent met the spec. After every main milestone and even every operate, run assessments or at the very least do fast handbook checks. If one thing fails, replace the spec or immediate earlier than continuing. For instance, if the spec stated, “Passwords have to be hashed with bcrypt” and also you see the agent’s code storing plain textual content, cease and proper it (and remind the spec or immediate in regards to the rule). Automated assessments shine right here: When you offered assessments (or write them as you go), let the agent run them. In lots of coding agent setups, you possibly can have an agent run npm check or related after ending a job. The outcomes (failures) can then feed again into the following immediate, successfully telling the agent “Your output didn’t meet spec on X, Y, Z—repair it.” This sort of agentic loop (code > check > repair > repeat) is extraordinarily highly effective and is how instruments like Claude Code or Copilot Labs are evolving to deal with bigger duties. At all times outline what “carried out” means (by way of assessments or standards) and test for it.

Iterate on the spec itself: When you uncover that the spec was incomplete or unclear (possibly the agent misunderstood one thing otherwise you realized you missed a requirement), replace the spec doc. Then explicitly resync the agent with the brand new spec: “I’ve up to date the spec as follows… Given the up to date spec, regulate the plan or refactor the code accordingly.” This manner the spec stays the only supply of fact. It’s just like how we deal with altering necessities in regular dev, however on this case you’re additionally the product supervisor to your AI agent. Preserve model historical past if attainable (even simply by way of commit messages or notes), so you understand what modified and why.

Make the most of context administration and reminiscence instruments: There’s a rising ecosystem of instruments to assist handle AI agent context and information. As an example, retrieval-augmented era (RAG) is a sample the place the agent can pull in related chunks of information from a information base (like a vector database) on the fly. In case your spec is large, you possibly can embed sections of it and let the agent retrieve probably the most related elements when wanted, as a substitute of at all times offering the entire thing. There are additionally frameworks implementing the Mannequin Context Protocol (MCP), which automates feeding the correct context to the mannequin primarily based on the present job. One instance is Context7 (context7.com), which might auto-fetch related context snippets from docs primarily based on what you’re engaged on. In follow, this may imply the agent notices you’re engaged on “cost processing” and it pulls the funds part of your spec or documentation into the immediate. Take into account leveraging such instruments or organising a rudimentary model (even a easy search in your spec doc).

Parallelize rigorously: Some builders run a number of agent cases in parallel on totally different duties (as talked about earlier with subagents). This will velocity up growth (e.g., one agent generates code whereas one other concurrently writes assessments, or two options are constructed concurrently). When you go this route, make sure the duties are actually unbiased or clearly separated to keep away from conflicts. (The spec ought to word any dependencies.) For instance, don’t have two brokers writing to the identical file directly. One workflow is to have an agent generate code and one other evaluation it in parallel, or to have separate parts constructed that combine later. That is superior utilization and may be mentally taxing to handle. (As Willison admitted, working a number of brokers is surprisingly efficient, if mentally exhausting!) Begin with at most 2–3 brokers to maintain issues manageable.

Model management and spec locks: Use Git or your model management of selection to trace what the agent does. Good model management habits matter much more with AI help. Commit the spec file itself to the repo. This not solely preserves historical past, however the agent may even use git diff or blame to grasp modifications. (LLMs are fairly able to studying diffs.) Some superior agent setups let the agent question the VCS historical past to see when one thing was launched—surprisingly, fashions may be “fiercely competent at Git.” By maintaining your spec within the repo, you permit each you and the AI to trace evolution. There are instruments (like GitHub Spec Package talked about earlier) that combine spec-driven growth into the Git workflow—as an illustration, gating merges on up to date specs or producing checklists from spec objects. Whilst you don’t want these instruments to succeed, the takeaway is to deal with the spec like code: Keep it diligently.

Price and velocity issues: Working with giant fashions and lengthy contexts may be sluggish and costly. A sensible tip is to make use of mannequin choice and batching neatly. Maybe use a less expensive/quicker mannequin for preliminary drafts or repetitions, and reserve probably the most succesful (and costly) mannequin for closing outputs or complicated reasoning. Some builders use GPT-4 or Claude for planning and significant steps, however offload less complicated expansions or refactors to a neighborhood mannequin or a smaller API mannequin. If utilizing a number of brokers, possibly not all have to be high tier; a test-running agent or a linter agent may very well be a smaller mannequin. Additionally take into account throttling context dimension: Don’t feed 20K tokens if 5K will do. As we mentioned, extra tokens can imply diminishing returns.

Monitor and log the whole lot: In complicated agent workflows, logging the agent’s actions and outputs is crucial. Verify the logs to see if the agent is deviating or encountering errors. Many frameworks present hint logs or permit printing the agent’s chain of thought (particularly in the event you immediate it to assume step-by-step). Reviewing these logs can spotlight the place the spec or directions may need been misinterpreted. It’s not in contrast to debugging a program—besides the “program” is the dialog/immediate chain. If one thing bizarre occurs, return to the spec/directions to see if there was ambiguity.

Be taught and enhance: Lastly, deal with every mission as a studying alternative to refine your spec-writing talent. Possibly you’ll uncover {that a} sure phrasing persistently confuses the AI, or that organizing spec sections in a sure means yields higher adherence. Incorporate these classes into the following spec. The sphere of AI brokers is quickly evolving, so new finest practices (and instruments) emerge always. Keep up to date by way of blogs (like those by Simon Willison, Andrej Karpathy, and so forth.), and don’t hesitate to experiment.

A spec for an AI agent isn’t “write as soon as, carried out.” It’s a part of a steady cycle of instructing, verifying, and refining. The payoff for this diligence is substantial: By catching points early and maintaining the agent aligned, you keep away from pricey rewrites or failures later. As one AI engineer quipped, utilizing these practices can really feel like having “a military of interns” working for you, however it’s important to handle them properly. A great spec, constantly maintained, is your administration device.

Keep away from Widespread Pitfalls

Earlier than wrapping up, it’s value calling out antipatterns that may derail even well-intentioned spec-driven workflows. The GitHub examine of two,500+ agent information revealed a stark divide: “Most agent information fail as a result of they’re too obscure.” Listed here are the errors to keep away from:

Imprecise prompts: “Construct me one thing cool” or “Make it work higher” provides the agent nothing to anchor on. As Baptiste Studer places it: “Imprecise prompts imply flawed outcomes.” Be particular about inputs, outputs, and constraints. “You’re a useful coding assistant” doesn’t work. “You’re a check engineer who writes assessments for React parts, follows these examples, and by no means modifies supply code” does.

Overlong contexts with out summarization: Dumping 50 pages of documentation right into a immediate and hoping the mannequin figures it out hardly ever works. Use hierarchical summaries (as mentioned in precept 3) or RAG to floor solely what’s related. Context size shouldn’t be an alternative to context high quality.

Skipping human evaluation: Willison has a private rule—“I gained’t commit code I couldn’t clarify to another person.” Simply because the agent produced one thing that passes assessments doesn’t imply it’s right, safe, or maintainable. At all times evaluation crucial code paths. The “home of playing cards” metaphor applies: AI-generated code can look strong however collapse below edge instances you didn’t check.

Conflating vibe coding with manufacturing engineering: Speedy prototyping with AI (“vibe coding”) is nice for exploration and throwaway tasks. However delivery that code to manufacturing with out rigorous specs, assessments, and evaluation is asking for hassle. I distinguish “vibe coding” from “AI-assisted engineering”—the latter requires the self-discipline this information describes. Know which mode you’re in.

Ignoring the “deadly trifecta”: Willison warns of three properties that make AI brokers harmful: velocity (they work quicker than you possibly can evaluation), nondeterminism (similar enter, totally different outputs), and value (encouraging nook reducing on verification). Your spec and evaluation course of should account for all three. Don’t let velocity outpace your skill to confirm.

Lacking the six core areas: In case your spec doesn’t cowl instructions, testing, mission construction, code fashion, git workflow, and bounds, you’re probably lacking one thing the agent wants. Use the six-area guidelines from part 2 as a sanity test earlier than handing off to the agent.

Conclusion

Writing an efficient spec for AI coding brokers requires strong software program engineering ideas mixed with adaptation to LLM quirks. Begin with readability of function and let the AI assist broaden the plan. Construction the spec like a critical design doc, protecting the six core areas and integrating it into your toolchain so it turns into an executable artifact, not simply prose. Preserve the agent’s focus tight by feeding it one piece of the puzzle at a time (and take into account intelligent ways like abstract TOCs, subagents, or parallel orchestration to deal with large specs). Anticipate pitfalls by together with three-tier boundaries (at all times/ask first/by no means), self-checks, and conformance assessments—primarily, train the AI how one can not fail. And deal with the entire course of as iterative: use assessments and suggestions to refine each the spec and the code constantly.

Comply with these pointers and your AI agent can be far much less more likely to “break down” below giant contexts or wander away into nonsense.

Joyful spec-writing!