- In short
- A goal-based prompt tells a subagent what to achieve and how its output will be judged, leaving the method to the model. A step-based prompt dictates the exact procedure to follow. Goal-based prompts let a subagent adapt to what it discovers, which is usually what complex research tasks need; step-based prompts trade that adaptability for control.
Goal based vs step based prompts: the core distinction
The choice between goal based vs step based prompts is really a choice about who decides the method. A goal-based prompt hands the subagent an objective and the standard its output will be held to, then gets out of the way: find the three strongest competitors and justify the ranking with evidence. A step-based prompt instead scripts the path: open this page, copy this table, then summarise rows one to ten. Both can work, but they fail in opposite ways, and analysing that trade-off is what this knowledge point asks of you.
The reason the distinction carries weight in agentic design is that a subagent runs alone, in a fresh context, and reports back only when done. You cannot steer it mid-task. So the prompt is not a suggestion, it is the entire operating contract. If you wrote rigid steps and the world does not match them, the subagent has no licence to recover. If you wrote a clear goal, it can route around the surprise. That is why Anthropic guidance favours focused, outcome-oriented instructions for workers handling anything open-ended.
- Goal-based vs step-based prompts
- Goal-based prompts specify the desired outcome and quality criteria and leave the method to the model. Step-based prompts specify the exact procedure to follow. The first maximises subagent adaptability; the second maximises control at the cost of flexibility.
What a goal based prompt looks like
A strong goal-based prompt is precise where it counts and open where it should be. It names the objective unambiguously, states the criteria by which success is measured, and supplies the context and constraints the subagent must respect, while deliberately leaving the procedure unspecified. Goal-based does not mean hand-wavy. Telling a subagent to "research competitors" is underspecified and will produce thin work; telling it to "identify the three competitors most likely to win enterprise deals and support each with pricing and feature evidence" is a goal with a quality bar.
This is also why goal-based prompts pair so naturally with structured context passing. When the objective is clear and the supporting findings arrive with their sources attached, the subagent has everything it needs to choose a sensible method and produce attributable output. The openness is about the how, never about the what or the how-well.
What a step based prompt looks like
A step-based prompt is procedural: do this, then this, then that. It shines when the procedure genuinely is fixed, when an audit trail matters, or when a wrong improvisation would be costly, for example a deployment checklist or a compliance routine. In those settings the rigidity is a feature, because you want the same path every time and no creative detours.
The cost shows up the moment reality diverges from the script. A step-based subagent told to read rows one to ten cannot react when the data it needs is in rows eleven to twenty, because adapting was never in its remit. The exam trap here is assuming that more prescriptive always means more reliable. For predictable, bounded jobs that holds; for exploratory work it inverts, because the steps encode assumptions that the task quietly violates.
Reading the trade-off like an architect
The analyse-level skill this knowledge point tests is matching prompt style to task shape. Ask two questions of any delegated job. First, can the steps be known in advance and will they stay valid? If yes, a step-based prompt buys you control cheaply. Second, will the subagent encounter things you cannot anticipate? If yes, a step-based prompt becomes a liability and a goal-based prompt with strong criteria is the safer design. Most genuinely agentic research tasks land in the second bucket, which is why the default lean is goal-based.
Hybrid prompts: goals with guardrails
The two styles are endpoints of a spectrum, not a binary, and the most robust real prompts borrow from both. A hybrid prompt states the goal and the quality criteria, then pins down the handful of constraints that are genuinely non-negotiable, and leaves everything else open. You might tell a subagent to surface the strongest security weaknesses and rank them by severity, the goal, while also insisting it never modifies files and always reports its evidence, the guardrails. The subagent keeps its freedom to investigate while staying inside the boundaries that actually matter.
The art is deciding which constraints earn a place. A constraint belongs in the prompt when violating it would be unsafe, non-compliant, or simply wrong regardless of what the subagent discovers. A constraint does not belong when it merely encodes one plausible method among many, because that quietly turns a goal-based prompt back into a brittle step-based one. The analyse-level skill is separating the few hard rails from the many soft preferences, and only hard-coding the rails. Get that separation right and you keep adaptability where it helps while removing it precisely where it would be dangerous.
Symptoms that your prompt style is wrong
You can often diagnose a mismatched prompt style from the output alone, which is exactly what an exam scenario asks you to do. Over-prescription shows up as a subagent that completes its instructions perfectly yet misses the obvious, because the steps never pointed it at the real issue and it had no licence to look elsewhere. The work is technically compliant and substantively useless, a tell-tale signature of a step-based prompt aimed at an open-ended task.
Under-specification shows up as the opposite: a subagent that wanders, produces broad and shallow output, or solves a different problem than you intended, because the goal and its quality bar were never made explicit. Here the cure is not to bolt on rigid steps but to sharpen the objective and the success criteria. Reading these two signatures, compliant-but-blind versus busy-but-unfocused, lets you prescribe the right fix: loosen toward goals when the task is exploratory, tighten the objective when the work is drifting. That diagnostic move is the heart of what this knowledge point assesses.
Over-constraining a goal-based prompt
There is a quieter failure that catches prompts which are otherwise goal-based: over-constraining the form of the answer. A prompt can name the outcome and the quality bar correctly and still smuggle in a rigid format requirement that hands back the very adaptability the goal was meant to preserve. Demanding a fixed table layout, an exact section order, or a precise word count turns a free-method instruction into a constrained one through the back door, and the subagent ends up shaping its findings to fit the mould rather than reporting what the task actually warranted.
The most common version is conflating length with detail. Telling a subagent to produce "a 1,000-word analysis" measures effort by volume, so a worker that finds a crisp, decisive answer pads it to reach the count, while a worker facing a genuinely complex case truncates to stay under it. The better move is to state the substantive bar, cover these competitors with pricing evidence and a ranked recommendation, and let the length follow the content. Format belongs in a goal-based prompt only when the downstream consumer truly needs a specific shape, such as a JSON record another stage will parse; otherwise an open format keeps the worker free to match its output to what it found.
Diagnosing this on the exam means separating two kinds of over-constraint. Procedural over-prescription dictates the steps and blocks adaptation to surprises, the classic step-based trap. Format over-constraint dictates the shape of the result and quietly distorts the content to fit it. Both narrow a goal-based prompt back toward rigidity, and the architect's move is the same in spirit: keep the constraints the outcome genuinely requires and drop the ones that only encode a preference about how the answer should look.
How this is tested on the exam
Domain 1 questions on this knowledge point are comparative. A scenario describes a subagent that performs poorly and gives you its prompt, and you must judge whether the prompt style is the problem. The classic failure is an over-prescriptive step-by-step prompt for a research task: the subagent finds something unexpected and cannot use it because the steps never allowed for it. The correct analysis is that the rigid procedure prevented adaptation, and the fix is to restate the prompt around the goal and the quality criteria.
Beware the inverse distractor too: a scenario where the procedure really is fixed and safety-critical, where switching to a loose goal would remove necessary control. The exam wants you to choose by task shape, not by slogan. Goal-based is the usual default for open-ended work, but step-based is correct when predictability and auditability are the point.
Worked example
A coordinator delegates a security audit of a web service to a subagent. The subagent keeps missing real issues, and its prompt is a fixed ten-step checklist.
The original prompt walks the subagent through ten exact actions: check these five headers, look for these three known CVEs, confirm TLS, and report. It is tidy and easy to review. But the service under audit has an authentication flaw that none of the ten steps touch, so the subagent dutifully completes the checklist, finds nothing in those ten places, and reports a clean bill of health while the real vulnerability sits untouched.
The procedure encoded an assumption, that the only risks are the ten the author thought of, and the task violated it. Because the prompt was purely step-based, the subagent had no mandate to investigate anything off-script, even when a smell was right in front of it.
Rewriting the prompt around the goal changes the outcome. The new prompt sets the objective, surface and prioritise security weaknesses in this service, and the quality bar, justify each finding with evidence and rank by severity, while offering the ten checks as a starting point rather than the whole job. Now the subagent runs the obvious checks and follows the authentication smell it noticed, because adapting to findings is exactly what the goal licenses. Same subagent, same model, different prompt shape, materially better audit.
Common misreadings to avoid
Misconception
Step-by-step instructions are always more reliable than goal-based ones for subagents.
What's actually true
Misconception
A goal-based prompt is just a vague prompt that hopes the model figures it out.
What's actually true
Putting it to work
Once you can choose a prompt style deliberately, the rest of task statement 1.3 gets easier. The same goal-with-criteria framing keeps parallel subagent spawning coherent when several workers run at once, and it complements fork session exploration, where each branch pursues the same goal down a different path. Prompt style is a design lever, and analysing when to pull it is a foundations-level architect skill.
A useful way to lock this in is to remember that the prompt encodes your trust. A step-based prompt says you trust your own foresight more than the subagent judgement, so you spell out the path. A goal-based prompt says you trust the subagent judgement more than your ability to predict the terrain, so you state the destination and the standard and let it navigate. Neither stance is universally right; the correct one depends entirely on whether the terrain is knowable in advance. Architects who pick the prompt style on purpose, naming why they trust foresight here and judgement there, design multi-agent systems that hold up under the surprises that real tasks always produce, and that deliberate reasoning is precisely what a comparative exam question is built to reward.
A subagent is asked to research emerging risks in a new market but consistently returns shallow, off-target results. Its prompt is a strict eight-step procedure that lists exactly which pages to read and in what order. What is the best diagnosis and fix?
People also ask
What is the difference between a goal based and a step based prompt?
When should I use a step based prompt for a subagent?
Does goal based mean vague?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
Designing effective subagents
Why watch: Specifying goals and quality criteria rather than rigid steps is the goal-based prompt design this KP champions.
More videos for this concept
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.