Elena Vasquez

Breaks complex requirements into sequenced, testable prompt steps with clear constraints and expected outputs for each stage.

Elena consistently decomposes feature requests into 3–5 focused prompts, each scoped to a single responsibility. Her decomposition of the payment flow into validation, processing, and notification steps was exemplary.

Predictable AI output with minimal rework. Each step is small enough to verify independently, keeping the feedback loop tight.

Start adding explicit failure-mode prompts for edge cases — decompose not just the happy path, but anticipated error states.

Specifies explicit boundaries — input types, output formats, forbidden patterns, performance targets — before prompting, so AI operates within a well-defined box.

Elena defines type constraints and output formats consistently, but sometimes omits performance bounds and edge-case restrictions. The search endpoint prompt lacked pagination limits.

AI generates code that fits the architecture but occasionally introduces unbounded behaviors that surface only under load or unusual inputs.

Add a "constraints checklist" to your prompt template: types, formats, error handling, performance bounds, and forbidden patterns.

Treats every AI output as a hypothesis. Runs tests, checks edge cases, reads generated code line-by-line, and validates assumptions before merging.

Elena runs the happy path but tends to accept plausible-looking output without verifying edge cases or reading through generated utility functions. Two bugs in Q1 came from untested null-handling in AI-generated code.

Occasional production issues from unverified assumptions. The team spends extra review cycles catching what could have been caught at generation time.

Before accepting any AI output, ask: "What would break this?" Write one negative test case for every positive one.

Structures prompts and codebases so the AI only sees the minimal relevant context — no leaking of unrelated modules, configs, or state.

Elena uses hexagonal boundaries effectively for domain logic but occasionally feeds full file contents rather than extracted interfaces when prompting for adapter code.

Generally clean outputs, but adapter implementations sometimes reference internal domain details they should not know about.

Create interface-only context snippets for cross-boundary prompts. Never pass implementation files when asking AI to write against a contract.

Reviews AI-generated code for reasoning quality, not just syntax. Challenges logical assumptions, checks for hallucinated APIs, and validates architectural fit.

Elena catches syntax issues and obvious bugs but does not consistently challenge the reasoning behind AI design choices. Accepted a caching strategy that looked correct but used an inappropriate eviction policy.

Code that passes linting and tests but may carry subtle architectural debt. Reasoning gaps accumulate over multiple AI-assisted iterations.

For every AI-generated function over 20 lines, write a one-sentence justification of its design choice before approving.

Writes executable specifications before prompting AI. Uses tests as constraints that guide generation, not just as post-hoc validation.

Elena starts most features with failing tests that define the expected behavior. Her TDD cadence — red, prompt, green, refactor — is consistent and well-internalized.

High-confidence AI output that is immediately verifiable. Significantly lower rework rates compared to team average.

Extend test-first thinking to integration boundaries — write contract tests for service interactions before prompting AI for implementations.

Designs multi-step AI workflows with clear boundaries, fallback strategies, and human checkpoints. Knows when to chain prompts vs. when to stop and verify.

Elena has experimented with chained prompt workflows but does not yet define explicit checkpoints or fallback strategies. Her agent pipeline for the data migration ran end-to-end without intermediate validation.

Risk of cascading errors in multi-step workflows. A single hallucination early in the chain propagates through all downstream steps.

Add a verification gate after every second step in any chain longer than three prompts. Define what "abort and retry" looks like for each stage.