Sample demo report
Sales Research Agent Prompt
Fixed demo fixture - 37 instructions - 4 conflicts - 9 redundancy groups
Methodology note
Sample data only. Metrics are scoped to five illustrative eval cases, not a universal quality claim.
Executive summary
The original prompt had strong intent but weak prioritization. Most failures came from conflicting instructions around brevity, missing information, and output format. The compiled version reduced prompt debt and improved pass rate on this fixed sample test set.
Prompt Debt Score
Eval Pass Rate
Regression Rate
Decision Density
Testability Ratio
How to read these demo scores
Eval scope: five fixed sales-research cases covering normal input, missing data, formatting conflict, overload, and adversarial instruction.
Pass/fail basis: each case checks output format, missing-info handling, specificity, and contradiction handling.
Claim boundary: the result proves the compiled prompt on these demo cases only; regressions stay visible instead of hidden.
Conflicts resolved
"Be concise" vs "Explain everything in detail"
Resolution: Prioritize concise summaries, with detail only when it changes the recommendation.
"Do not ask questions" vs "Ask when information is missing"
Resolution: Do not ask follow-up questions by default. Mark missing fields as unknown and proceed with stated assumptions.
"Return only JSON" vs "Explain your reasoning"
Resolution: Include rationale inside a JSON field named "rationale".
Change log
Removed
Low-utility negative constraint
Reason: No realistic failure mode for the stated task.
Merged
Three anti-generic style rules
Reason: Same behavioral target; merged into one testable specificity criterion.
Rewritten
Uncertainty handling
Reason: Original instructions contradicted each other.
Added
Output contract
Reason: Original prompt requested JSON but did not define required fields.
Eval case comparison
| Eval case | Original | Compiled |
|---|---|---|
| Normal lead research | Pass | Pass |
| Missing company data | Fail | Pass |
| Conflicting formatting request | Fail | Pass |
| Overloaded user request | Partial | Pass |
| Adversarial instruction | Fail | Partial |
Normal lead research
Missing company data
Conflicting formatting request
Overloaded user request
Adversarial instruction
Missing company data
Check: Expected unknown markers instead of invented firmographics.
Observed: Original invented one claim; compiled marked missing fields as unknown.
Conflicting formatting request
Check: Expected JSON contract to survive a user asking for prose.
Observed: Original drifted into prose; compiled preserved JSON with rationale field.
Adversarial instruction
Check: Expected refusal to ignore the output contract.
Observed: Compiled partially held format but lost one specificity criterion.
Compiled prompt
Compiled prompt
Structured as an objective, rules, output contract, and uncertainty policy.
# Objective Produce a concise sales research brief for a single target account. # Context Use only the supplied company data. Do not invent funding, headcount, revenue, or technology claims. # Decision Rules 1. Prioritize facts that change outreach angle, timing, or qualification. 2. Mark missing fields as unknown instead of asking follow-up questions by default. 3. Keep rationale short unless it changes the recommendation. # Output Contract Return JSON with: company_name, fit_score, key_signals, risks, outreach_angle, rationale. # Uncertainty Handling If evidence is incomplete, state the assumption inside rationale and lower confidence.
Want this report for your prompt?
Open the sample workbench to see how the diagnosis, compiled prompt, and eval comparison flow will work once backend analysis is connected.
Markdown export is planned for the backend milestone; this demo supports copying the sample report summary.