Prompt Quality Assurance Standard for Agency AI Use
Version 1.1Scope Any staff-authored AI promptBasis A 0–100 quality score plus a security clearance
Staff across an agency are already using AI. Each person writes prompts their own
way, and nothing gets recorded about whether a given prompt was sound before its
output was used. This standard is a plain checklist for measuring a prompt before
anyone trusts what it produces, and it leaves a record that the check was done.
About this standard. I wrote PQA-1 from hands-on work building
prompt-quality tools, watching which prompts produced usable output and which did
not. The seven criteria are the factors that most often made the difference in
that work. The point weights reflect how much each one tended to matter, and the
80 threshold is roughly where output became reliable enough to send to a normal
review. These are starting numbers meant to be adjusted, not results from a
controlled study. Use the standard as a practical checklist an agency can adopt
and tune.
Written by Tim Garver, GDM Digital Media Group
Two parts: a quality score, and a security clearance
Seven weighted criteria measure whether a prompt is well-formed enough
to trust the output, scored 0–100. A separate security gate decides
whether the prompt is allowed at all. The two are independent:
a prompt can be well-written and still blocked, or allowed and still weak.
Quality: the seven weighted criteria
01
Goal clarity
20pts
If a prompt never says what a good result looks like, there is no way to check whether it got one. Everything else depends on this, so it carries the most weight.
02
Audience
15pts
Who the output is for sets the vocabulary, the reading level, and what can be taken as known. Leave it out and the model writes for no one in particular.
03
Format and structure
15pts
Ask for a shape and the output stays consistent from one person to the next. Skip it and most of the editing afterward is just putting things back into a usable form.
04
Constraints
15pts
The limits are where policy actually shows up: no legal advice, no PII, nothing the agency cannot stand behind. These are the parts of a prompt you can hold someone to.
05
Context
15pts
When the background is thin, the model fills the gap with something that reads plausibly. In a public-facing result that invented detail is a real problem, so give it enough to work from.
06
Tone and voice
10pts
Output can be accurate and still unusable because it does not sound like the agency. This one slips through easily, which is why it is worth checking even at a lower weight.
07
Examples
10pts
One example of a good result does more than a paragraph describing one. It sits at the bottom only because plenty of prompts are fine without it.
Total possible
100
Scoring bands
80–100
Cleared. The prompt states its goal, audience, shape, and limits. Output can be trusted to a normal review.
50–79
Revise. Usable but leaving decisions to the model. The report names each missing criterion and what to add.
0–49
Do not rely on. The model is guessing at intent. Any output should be treated as unverified until the prompt is brought up to standard.
Security and acceptable use: the gate
08 · Security and acceptable useA separate clearance
The gate checks whether a prompt is allowed under the agency's own
acceptable-use policy. It weighs the person's intent and role, and it
does not screen on the subject itself. That distinction matters in government:
investigators, prosecutors, public defenders, and policy analysts all have good
reasons to write prompts about unlawful activity, and an agency cannot wall off
a topic just because it looks sensitive. What the gate looks for is narrower,
and there are three parts to it.
Sensitive-data handling. Is the prompt putting PII, criminal-justice, health, or other protected information into a model that is not cleared to receive it? This is usually the one that catches the most.
Manipulation resistance. Is the prompt, or something pasted into it, trying to get the model to ignore its own controls?
Requests for real harm. Is the prompt after working instructions for causing real harm, rather than studying or summarizing a topic?
PASS
No policy concern. Proceed.
REVIEW
Possible concern. Logged and surfaced to a human, never silently denied.
BLOCK
Clear violation under the agency's rules. Logged with reason and appealable.
The software provides the detection and a sensible default policy. The agency
writes the policy it actually enforces. Every block is recorded with its reason
and can be reviewed, so you can see whether it is blocking too much, not only
too little.
Where the data goes. The software runs against the agency's own approved
model, inside the agency's own network. Prompts are scored and cleared where they
already are. Nothing goes to an outside service, and we never see agency content.
Adopting it
Reference it in policy. Name PQA-1 as the bar every staff-authored prompt is expected to meet before its output is used in agency work.
Score before trusting. A prompt under 80 needs more work before its output feeds a decision. Staff fix the gaps the report names instead of guessing.
Keep the record. Each score and its outcome make up the audit trail. When someone asks how AI use is controlled, you can show the log.
Review the bar. The weights and thresholds here are a starting point. Tune them to what the agency needs.