Thriving as an AI-Assisted Engineer

Define measurable acceptance criteria for quality, speed, cost, and reliability.

Construct an assumption map that separates facts, hypotheses, and unknowns.

Prioritize assumptions by impact and strength of existing evidence.

Design the smallest experiment that can test the highest-risk assumption.

Identify the decision owner, contributors, reviewers, and affected stakeholders for a task.

Translate informal stakeholder requests into explicit scope and non-goals.

Formulate clarification questions that expose hidden constraints and disagreements.

Define ownership and handoff boundaries for a cross-functional task.

Write a concise decision brief that records evidence, decisions, risks, and open questions.

Allocate task steps between yourself, Claude Code, and deterministic tooling according to failure risk.

Assemble a compact context package containing the repository facts and constraints needed for a coding task.

Elicit an implementation plan that states assumptions, affected components, and planned validation.

Critique a Claude Code implementation plan before authorizing code changes.

Partition an implementation plan into independently reviewable increments.

Create explicit checkpoints that prevent an agent session from proceeding on unverified assumptions.

Rank the behavioral, data, security, and operational risks in a proposed Python change.

Trace one critical execution path through an unfamiliar generated code change.

Derive tests from requirements without copying the generated implementation's assumptions.

Express behavioral invariants as property-based tests.

Construct negative tests for boundary conditions and dependency failures.

Verify generated claims about libraries and APIs against primary sources.

Decide whether a generated change is acceptable using a documented evidence threshold.

Translate a business decision into an analytical question with a defined population, unit of analysis, and metric.

Audit a data source for provenance, coverage, selection bias, and measurement error.

Design a sampling strategy appropriate to an exploratory research task.

Quantify uncertainty in an analytical result using sensitivity analysis.

Triangulate a claim using signals with different biases.

Create a reproducible evidence table that links each conclusion to its source and caveats.

Decompose an LLM agent workflow into stages with explicit input and output contracts.

Require structured outputs that preserve provenance for agent-generated claims.

Construct a small evaluation set containing representative, boundary, and adversarial cases.

Define a scoring rubric for correctness, completeness, traceability, and usefulness.

Measure agent quality by failure category rather than relying on an aggregate score alone.

Add abstention, retry, fallback, or human-review behavior for a specified failure mode.

Select a thin vertical slice that tests the highest-risk part of a proposed system.

Define an evidence package that must accompany a pull request or analytical deliverable.

Specify observability signals for quality, failures, latency, and cost.

Design a staged rollout with explicit rollback triggers.

Allocate verification effort according to impact, reversibility, and uncertainty.

Run a delivery retrospective using cycle time, rework, escaped defects, and agent mistakes to select the next deliberate-practice target.