Prompt Scoring | Smallest AI Docs

Seeing Your Score

On the Prompt tab of a single-prompt agent, click Review at the bottom of the prompt area to score the prompt. Once it’s scored, the bar shows Strong prompt (overall grade Excellent or Good) or Weak prompt (Needs Work or Poor).

Weak prompt badge with View N suggestions link — Weak prompt badge in the agent editor. Click the link next to it to open the analysis panel.

Click View N suggestions (or View analysis when there are none) to open the analysis panel. If you edit the prompt after scoring, run the review again to refresh the score.

The Analysis Panel

The panel shows your prompt on the left and the findings on the right. Each finding card names a check, suggests a fix, and shows its level: Needs attention, Can improve or Looks good. Use the filter buttons above the cards (All, Needs attention, Can improve, Looks good) to narrow the list.

Prompt analysis panel showing the overall score, latency estimate, and per-dimension findings.

Panel field	What it means
Overall grade	Numeric score (0–100) and label - Excellent / Good / Needs Work / Poor
Estimated latency	Estimated time-to-first-token overhead added by your prompt length
Weak prompt / Strong prompt	Summary label derived from the overall score
Low / Normal / High latency	Token-density band for the current prompt

How Scoring Works

Each time you request a score, Atoms sends your prompt through two sequential Gemini passes - a Platform Analyst pass and a Rubric Judge pass - and returns scores for 11 dimensions grouped into three priority tiers.

Tiers and dimensions

Tier 1 - Core (highest priority)

Tier 2 - Quality

Tier 3 - Integrity (gating)

Dimension	What is checked
Role & Objective	Is the agent’s purpose and scope clearly defined?
Personality & Voice	Are tone and style specified with concrete guidance, not just adjectives?
Conversation Structure	Are there explicit flow phases and exit criteria?
Tool Integration	Are all referenced tools declared, with failure paths covered?
Constraints & Safety	Are safety rules, escalation paths, and hard limits spelled out?

Each dimension is rated Strong, Adequate, Weak, Missing, or Not Applicable.

Quality Levels and Token Bands

Grade thresholds

Score	Grade
90–100	Excellent
75–89	Good
50–74	Needs Work
0–49	Poor

Token density bands

The Estimated latency figure and the density band are derived from your prompt’s token count.

Band	Token range	Latency impact
Lean	Fewer than 4K tokens	Very low
Normal	4K–9.9K tokens	Low
Heavy	10K–14.9K tokens	Moderate
Overweight	15K or more tokens	High

For voice agents, aim for the Normal band. Heavy and Overweight prompts increase first-token latency, which makes responses feel slower to callers.

Scoring via API

You can trigger scoring programmatically against a published version or a draft. Each call deducts 1 credit. Re-submitting an unchanged prompt returns a 400 - retrieve the cached result via GET /agent/{id} instead.

# Score a published version
curl -X POST https://api.smallest.ai/atoms/v1/prompt-scoring/score \
  -H "Authorization: Bearer $SMALLEST_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"versionId": "6a1589b75e048394eb37bc47"}'
# Score a draft
curl -X POST https://api.smallest.ai/atoms/v1/prompt-scoring/score \
  -H "Authorization: Bearer $SMALLEST_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"draftId": "6a1589b75e048394eb37bc48"}'

The response includes overall_score, overall_grade, band, estimated_ttft_overhead_ms, and a dimensions array with one entry per scored dimension.

Prompt scoring is only supported for single-prompt agents. Conversational flow (workflow-graph) agents return a 400.

→ Full API reference

Prompt

Best practices for structuring your system prompt

Agent Versioning

Publish and manage agent versions safely