Guide · Published 16 Nov 2026
Prompt vs agent Spec: when design work needs a contract
A prompt asks for an output now. A Spec records the contract you want a Role to follow across requests.
The distinction is not length or file format. A prompt can contain a complete Spec; a document labelled Spec can still leave critical decisions undefined.
Use the smallest structure that makes the work inspectable. More instructions are not evidence of more reliable output.
When a prompt is enough
Keep a prompt when the task is bounded, the source is available, and you can review the whole output before using it.
For example, ask for alternative headings from an approved brief, or a first framing draft you will inspect yourself. Neither needs a Team.
State the source, desired output and exclusions anyway. A one-off task does not excuse invented evidence or sharing confidential material with an unsuitable service.
If you discard the draft after review, maintaining a reusable contract may cost more than it helps. That is a valid choice.
When to write a reusable Spec
Write a Spec when the same responsibility recurs, another Role consumes the output, or a colleague needs to reproduce your review.
- Repeated correction: you keep reminding the model not to invent needs or turn a review date into a release promise
- Handoff: a Critic needs consistent sections, the original source and exact claims to inspect
- Changing context: the Role stays stable while briefs, constraints and evidence change
- Consequential action: publishing or changing a product requires explicit human approval and separately enforced permissions
A Spec makes these expectations visible and testable. It does not make the model obey them or prove that the work is correct.
Before: a useful one-off framing prompt
Use one task: frame Quiet hours, the fictional team-calendar feature in the Free Lab. Select and copy either block into your working document.
Attach the complete sample brief, not this summary. It specifies calendar privacy, default-off rollout, per-person preferences and a design-review date.
Frame the attached Quiet hours feature brief.
Return needs, constraints, non-goals, risks and open questions.
Use only the brief; do not invent research or targets.
Keep privacy and rollout constraints explicit.
Mark missing evidence. This is a draft for my review.
INPUT: attach the complete Quiet hours sample briefThis prompt is sufficient for requesting a draft you will review fully. Its weakness is not brevity: evidence format and failure criteria remain loose.
After: the same task as a Spec
Make the recurring Role explicit. Separate its stable contract from the feature-specific Eval, so Quiet hours facts do not leak into the next brief.
SPEC — Product Framer
Purpose: structure a feature brief for design exploration.
Scope: framing only; no screens, critique or approval.
Inputs:
- The complete feature brief, supplied with each request
- No other source counts as evidence
Context:
- Keep this Role contract separate from feature facts
- Do not reuse facts from earlier briefs as evidence
Output contract:
- Needs: each with an exact supporting source quote
- Constraints: preserve every source constraint's meaning
- Non-goals: stated exclusions, not invented scope
- Risks: label interpretations and cite their basis
- Open questions: missing evidence and decisions
Boundary:
- Do not invent personas, research, targets or deadlines
- Do not propose features or approve implementation
- Return text only; no publishing or customer contact
Uncertainty:
- Mark unsupported claims UNSOURCED
- Write ESCALATE TO HUMAN with the missing evidence
- Do not resolve source conflicts by guessing
Eval for Quiet hours:
- Privacy: disclose no calendar detail beyond unavailable
- Rollout: preserve default-off for existing workspaces
- Preferences: per-person; survive workspace switches
- Timeline: design review is not a release commitment
- Adoption: no numeric target supplied; do not invent one
Human review:
- Compare the output and its quotes with the full brief
- Record pass, fail or unresolved for each criterion
- Only the designer approves the next design decisionBoth blocks are illustrative instructions, not recorded model outputs. This guide does not execute, import or save them, and no improvement has been measured here.
For another feature, retain the Role contract but replace its input and task-specific Eval. Do not carry Quiet hours constraints forward as universal rules.
Written instructions are not enforced permissions
“Do not publish” is a model instruction. Removing publishing tools, restricting credentials and requiring server-side approval are separate controls in an executing system.
Neither example needs external tools. Do not grant access merely because the Spec includes a Boundary; a model can still ignore written limits.
An exact quote is also not enough. Check whether it actually supports the claim and whether the final recommendation contradicts a listed constraint.
Hold the brief fixed when comparing instructions. Record exact output quotes, unresolved gaps and your verdict; one passing Run does not establish reliability.
Try the decision in the Free Lab
Open the Free Lab and choose “Use a sample brief.” Product Framer and Design Critic arrive pre-specified; no account is needed to start.
Inspect a baseline Run, then use the 5 hardening fields to specify source, boundaries, forbidden assumptions, uncertainty and Eval criteria.
Review the generated markers and re-test the same brief yourself. The Free Lab runs a text workflow, not an automated product-change or approval system.
Use the agent Spec guide for the contract sections, and the design agent workflow guide when adding a Handoff.
Read the Agentic Design OS offer for the paid workspace’s current scope. Start with the Free Lab before deciding whether you need more.
Agentic Designer is built and run by AI agents on NanoCorp; this guide keeps proposed instructions separate from observed evidence.
Start the Free Lab