The 7-Point Prompt Grading Rubric: Turn Prompt Iteration From Vibe to Measurable Comparison

**Dimension 1 — Specificity (1–5).** Does the output address THIS specific case, or could it have been written for any similar input? Score 5 if the output uses concrete details from the input (audience, context, examples). Score 1 if the output is generic — would fit any input in the category. Most weak LLM outputs score 2–3 here; the lift from this dimension alone is large.

**Dimension 2 — Audience-appropriateness (1–5).** Does the language, examples, and level of complexity match the defined audience? Score 5 if a member of the target audience would feel addressed precisely. Score 1 if the output assumes the wrong expertise level or uses inappropriate references. Common failure: writing for 'beginners' but using jargon a beginner wouldn't recognize.

**Dimension 3 — Format adherence (1–5).** Does the output match the requested structure (length, sections, bullet vs. paragraph, headings, tone)? Score 5 if format is exactly what was specified. Score 1 if the output ignores format requirements entirely. The cheapest dimension to fix; LLMs respond well to specific format requirements when stated clearly.

**Dimension 4 — Constraint compliance (1–5).** Does the output respect explicit constraints (forbidden words, required inclusions, word count, etc.)? Score 5 if every constraint is honored. Score 1 if multiple constraints are ignored. Specific constraints (avoid words X, Y, Z) work better than vague constraints (avoid clichés); your prompt's constraint quality affects this score.

**Dimension 5 — Coherence (1–5).** Does the output hang together — paragraphs flow, claims connect, sections support each other? Score 5 for tight logical flow. Score 1 for disconnected paragraphs that could be reordered without losing meaning. Coherence drops in long outputs; weak prompts produce weak coherence at length.

**Dimension 6 — Insight / non-obviousness (1–5).** Does the output say something that wouldn't have been in the first response a competent human would write? Score 5 if the output surfaces non-obvious connections or analysis. Score 1 if it's the most-common surface-level response. The hardest dimension to score consistently because it's somewhat subjective; useful to score against 'would a competent expert in this domain learn anything from this output?'

**Dimension 7 — Actionability (1–5).** If the output is meant to drive action (recommendations, advice, decisions), is the action clear and executable? Score 5 if the reader can act immediately on the output. Score 1 if the output describes a situation without pointing to specific next moves. Many LLM outputs score high on description and low on actionability; the gap is fixable in the prompt.

**Rule 1 — Score each dimension independently.** Don't let your overall impression bias individual scores. Read the output, score Dimension 1 (Specificity) without looking at others; then 2, etc. Independent scoring captures the texture of where the output is strong and weak.

**Rule 2 — Use the full 1–5 range.** Most graders cluster around 3–4 because '5 feels too strong.' Force the use of 5 (genuinely excellent on that dimension) and 1 (genuinely failed). Without using the full range, the rubric loses its discriminating power.

**Rule 3 — Anchor scores to examples.** Write down what a 5 looks like for each dimension before scoring. 'Specificity 5 = mentions the audience's specific pain point from the input by name.' 'Specificity 1 = could be reused for any input in the category.' Anchoring stops drift across iterations.

**Rule 4 — Score across multiple outputs at once, not one at a time.** Run 10 outputs from version A and 10 from version B. Score all 20 in one session against the rubric. Comparison signal is stronger than absolute scoring because you can see the dimensions where one version consistently beats the other.

**Rule 5 — Compute total + dimension-specific deltas.** Total score (out of 35) is your headline metric. But also track per-dimension averages across the 10 outputs — if Version B is +0.4 on Specificity but -0.6 on Format, you've gained one thing and lost another, and the next iteration should preserve the gain while fixing the loss.

**Step 1 — Establish baseline.** Run your current prompt on 10 representative inputs. Score all 10 outputs against the rubric. Compute average per-dimension score and total. This is your baseline.

**Step 2 — Identify the weakest dimension.** Look at the per-dimension averages. The lowest score is the biggest opportunity. Don't try to lift everything; pick the dimension where the gap to a 4 average is largest.

**Step 3 — Make ONE targeted change to address that dimension.** If weak dimension is Format, add a more specific format instruction. If weak dimension is Specificity, add explicit instruction to use specific details from the input. Don't change multiple things at once — you won't know what helped.

**Step 4 — Re-run + re-score.** Run the new prompt on the same 10 inputs (or 10 new ones). Score again. Compare per-dimension deltas. The targeted dimension should improve; other dimensions should be neutral. If they degraded, you made a trade-off; decide if the trade is acceptable.

**Step 5 — Iterate until total exceeds quality threshold or you hit diminishing returns.** Most prompts iterate 4–8 cycles before hitting their realistic ceiling. After 8 cycles with marginal improvement, the prompt is probably near its ceiling and further iteration is diminishing-return work. Stop iterating; ship.

**Pattern 1 — Halo effect from one strong dimension.** A response that's brilliantly specific tempts you to score every dimension high. Force yourself to score each dimension in isolation. A specifically brilliant response can still have weak coherence (Dimension 5) or weak actionability (Dimension 7); the rubric only works if you score them separately.

**Pattern 2 — Confirmation bias toward the version you wrote.** If you wrote the new prompt, you'll subconsciously want it to win. Mitigate by blinding scores when possible — have someone else label the outputs as 'A' or 'B' without telling you which version produced which, then score, then reveal.

**Pattern 3 — Scoring at the wrong altitude.** The rubric is for evaluating prompt quality, not output quality on dimensions the prompt didn't ask about. If the prompt didn't request actionability, don't score Dimension 7. Only score dimensions the prompt was actually shaping toward.

**Pattern 4 — Insufficient sample size.** Scoring 2 outputs gives you noise, not signal. 10 outputs is the minimum reliable sample size for noticing patterns. Below 10, individual output variance overwhelms the prompt-version variance you're trying to measure.

**Use this rubric when:** iterating on a prompt for content generation, long-form writing, advice/recommendation outputs, or any task where output quality is subjectively assessed. The 7 dimensions cover the bulk of what 'good' means for these tasks.

**Don't use this rubric for:** structured extraction with binary correct/incorrect outcomes (use accuracy metrics), classification (use precision/recall/F1), code generation (use compilation + functional tests), or any task where automated metrics are stronger than human rubric scoring.

**Hybrid approach for production systems:** automated metrics where they exist; this rubric for the human-judgment dimensions automated metrics don't capture. Many production LLM workflows benefit from both — accuracy/precision metrics for what's measurable; rubric scores for craft and audience-fit dimensions.

(Note: variants of this rubric exist in academic prompt-engineering literature; nothing here is original framework. The specific 7-dimension cut and the 1–5 scoring approach are what worked across the workflows I've iterated on; other configurations work too.)

Iterating by vibe ('the output reads better now'): no memory of what changed, no signal on which dimension to push next, no comparison signal across iterations. The default approach and the slowest path to a good prompt.
Iterating with the 7-point rubric: explicit scores per dimension, clear next-iteration target (lowest-scoring dimension), comparison signal across versions, faster convergence to the prompt's quality ceiling.