jevify
Home
Cases
HarnessJEV LabsConsole
Case archive

Harnesses people actually shipped

Every entry links to its published source. The code is quoted from that source and trimmed, never invented, and performance claims stay attributed to whoever made them.

Audit your own call-sites

prompts

Copy one, paste it into your agent.

These follow the same contract as the report this project generates: evidence per call-site, unknown left as unknown, no estimated savings.

01

Audit an entire project

Use this when: you need an evidence-based inventory before changing any AI call-site.

Paste into any agent. This is the same contract the jevify report follows.

copy-ready prompt
Audit this project for AI call-sites that ask a language model for a bounded decision.

For every call-site you find, report:
- file and line
- the evidence in the code that it is an AI invocation
- whether the output is open language or a bounded decision
- if bounded: choice (one of a known set), score (ordered scale) or noul (probability)
- the confidence you have, and what context is missing when you are not sure

Do not rewrite anything. Do not estimate savings, latency or cost.
If the output shape is not visible in the code, report the call-site as unknown.

02

Refactor one bounded decision

Use this when: Use this after choosing a specific call-site that should return a typed decision instead of prose.

copy-ready prompt
Refactor exactly one call-site: <file>:<line>.

Rules:
1. Keep the generative call if the output is language a person reads.
2. If the output is a bounded decision, replace it with a typed decision:
   state = the context the code already has, questions = the decision you need.
3. Branch on confidence. Below the threshold, escalate instead of guessing.
4. Never move arithmetic, permissions or side effects into the model.
5. Show the before and after, and the threshold you chose and why.

03

Separate judgment from generation

Use this when: one prompt both decides something and writes the final text a person reads.

copy-ready prompt
This call-site does two jobs in one prompt: it decides something and it writes something.

Split it:
- extract the decision into a typed question (choice / score / noul)
- keep generation for the text a human reads
- pass the decided value into the generation step as input, not as a request

Return the two calls and the code that connects them.

04

Run the audit through MCP

Use this when: Use this from an MCP client when you want jevify to inspect a repository or collected files directly.

Works from Lovable, Claude Code, Cursor or any MCP client that supports OAuth.

copy-ready prompt
Connect to the jevify MCP server at https://jevify.lovable.app/mcp and run:

1. audit_repository on <public github url> (or audit_files with the files you collected)
2. read the findings, then classify_ai_callsite on any call-site you disagree with
3. generate_jevify_report for the final report

Report the coverage counters with the result: files read, skipped, truncated.
Treat an unknown classification as unknown, not as a candidate.

Run the audit on your own code

Point the console at a public repository or paste the files. The report names every AI call-site it found, the evidence behind the classification, and the coverage it managed to read.

Open the console