Methodology

Prompts Are Engineering Artifacts.
Treat Them Like It.

Most prompts get written, copy-pasted, and forgotten. No tests. No version history. No way to compare two approaches. Promptvexity replaces that workflow with a structured one borrowed from how software actually gets built.

The Status Quo

Prompt engineering is mostly vibes.

You write a prompt. It works in your head. You ship it. It fails in production. You tweak a word. Maybe it works better. You can't tell.

There is no commit history. No diff between v1 and v2. No way to know if you actually improved the prompt or just got lucky on a different sample. Two engineers solving the same problem produce wildly different prompts and there is no shared yardstick to compare them.

# typical workflow
write prompt
try it
tweak a word
try it again
...
forget what you changed
forget what worked
copy-paste to Slack
never look at it again
The Loop

Problem. Prompt. Score. Improve.

A four-step cycle that makes prompts measurable. Each stage produces an artifact the next stage can reason about.

01 · Problem

Frame the outcome

Industry context, primary goal, example input, expected output, known failure modes, success criteria. Constraints are explicit, not implied.

02 · Prompt

Engineer the input

System prompt, user template with variables, model choice, parameters, usage context, tradeoffs. A prompt is a contract, not a sentence.

03 · Execute

Run it, keep the receipt

Prompts are executed against real models and every call is kept: the exact input, the output that came back, and what it cost. No single number standing in for whether something works.

04 · Improve

Iterate with lineage

Fork the prompt. Note what changed. Compare versions head-to-head on the same problem. The history is the artifact.

The Evidence

A number you cannot check is not evidence.

Every prompt carried a Quality Score out of 100. It stopped being calculated on 10 August 2026 after we checked it: the fifth unanimous upvote moved one prompt from 63 down to 56. What is left is smaller, and every part of it says plainly what it does and does not establish.

The same standard decided which problems exist at all. Read the falsification record, including the candidates that were dropped because no model failed them.

297 calls

Recorded runs

Prompts were run against real models and whatever came back is stored verbatim, including the 67 responses that came back empty and the 80 the model cut off at its token ceiling. The exact input sits next to the output, so you can send the same request from your own client and compare. It is a record of what happened, not a judgement of whether it was good.

• The exact system prompt and user message sent
• The output that came back, stored verbatim
• Input and output tokens, latency in ms, cost in USD
• 4 models, captured 8 August 2026, $0.296 total
• Public at /prompts/[slug]/transcripts
Readers

Community signal

Upvotes, downvotes, and Worked or Failed reports from people who say they used the prompt. The leaderboard ranks on a Wilson score lower bound of the votes, so an upvote can never move a prompt down it. The ranking is recomputed on a schedule, not on every vote.

• Upvotes and downvotes
• Worked and Failed reports on a prompt
• Wilson lower bound, implemented as public.wilson_lower_bound
• Monotonic by construction, unlike the score it replaced
• Self-reported, with nothing tying a report to a run
Not live

Declared tests

A table for declared test cases and assertions exists in the schema. It holds zero rows. Until it is filled, nothing here has been checked against a stated expectation.

• No pass and no fail on any of the 297 stored calls
• 292 read not judged; the 5 that errored read call errored
• eval_cases exists and is empty
• Per-problem suites with assertions are the next build
• We will say when it is live, and not before
Lineage

Every fork tells a story.

When you fork a prompt, the new version keeps a pointer to the parent. The fork summary explains what changed and why. The depth metric tracks how many iterations away from the root the prompt is.

A great prompt is rarely the first attempt. The lineage view shows the path: which version added an output format, which one tightened a constraint, which one finally beat the baseline.

v1.0
47
initial draft
v1.1
58
added JSON format spec
v1.2
71
added 3 worked examples
v2.0
88
tightened persona + edge cases
Best
Reputation

Authorship compounds.

Reputation is built from the quality of your prompts and the validation they earn. Tiers act as a credibility signal so readers know whose work has been battle-tested.

View the tier system

Pick a problem. Write a prompt. See the number.

The methodology is only useful if you run it. Start with a real problem from the catalog and submit your first scored prompt.