A.  

Crash on any unexpected output and require manual recovery on affected requests the application handles during normal operation.

B.  

Silently drop any output that does not match expectations so downstream systems do not see the malformed output the application receives from the model.

C.  

Use the output the model produces and leave schema enforcement to a later phase of production after the application is more complete.

D.  

Apply defensive parsing to handle unexpected fields, missing values, and type mismatches that the downstream code might otherwise crash on.

A.  

Switch to a different agentic framework that advertises automatic context-window management as a built-in feature.

B.  

Reduce the number of tool calls per task by combining several existing tools into larger, multi-purpose tools.

C.  

Increase the context window further so all tool outputs from every prior step remain in full detail throughout the task.

D.  

Apply a context-management pattern that summarizes or prunes older tool outputs while preserving the active task state.

A.  

Define the Skill as a long inline instruction at the top of every CLAUDE.md file in the team’s repositories so Claude has access to it on every session.

B.  

Define the Skill inside the application’s source code as a regular library module and call it from the application code instead of from Claude Code.

C.  

Define the Skill in a Skills directory recognized by Claude Code, where Claude can discover and load it during sessions for the team's repositories.

D.  

Define the Skill in a personal scratch directory on each developer's machine and load it manually before each Claude Code session that needs it.

A.  

Automated tests of the Claude integration, linting, and any other standard quality gates the team applies to its other services.

B.  

A full end-to-end production deployment on every pull request to catch all possible issues before any code is merged into the main branch.

C.  

Automated tests of the Claude integration only, with linting handled separately during local development on each developer's machine.

D.  

Automated linting and security scanning, with Claude integration testing handled manually during pre-release verification by a designated reviewer.

A.  

Duplicate CLAUDE.md and settings.json files in multiple repositories to provide redundancy, on the grounds that a single source of truth is risky for project configuration.

B.  

Version-control CLAUDE.md and settings.json files in a separate repository from the project's source code, so configuration evolves independently from the application code over time.

C.  

Version-control CLAUDE.md alongside the project's source code and settings.json files in a separate repository from the project's source code.

D.  

Version-control CLAUDE.md and settings.json files alongside the project's source code, with changes reviewed through standard pull request workflows the team applies.

A.  

Remove tool calling from the workflow entirely so the agent operates as a single text-generation step with no tool outputs accumulating in the context window.

B.  

Apply context engineering techniques such as tool output pruning or compaction to keep the active task state visible while reducing the volume of older content.

C.  

Increase the model's context window so the agent can hold every tool output at full detail across the entire workflow no matter how many tool calls it accumulates.

D.  

Restart the agent every five tool calls to prevent any drift, with the agent losing all task state at each restart point during the workflow.

A.  

Omit the tool description and let the model infer when to use the tool based on the tool's name and the rest of the prompt context.

B.  

Define the tool with a loose schema and let the model interpret the inputs flexibly on each call the agent makes.

C.  

Implement the tool with no error handling and let the agent loop catch failures whenever the pricing service returns an error during operation.

D.  

Define the tool with a clear schema, write a precise description for when to call it, and handle pricing service errors explicitly.

A.  

Add retry logic to each subagent so that when a tool call fails, the subagent retries automatically before escalating — and configure each subagent to log its tool calls and outputs to a shared trace so the team can audit redundancy and contradictions after the fact.

B.  

Give each subagent read access to the other subagents' outputs so they can identify and resolve contradictions without routing back through the primary agent.

C.  

Strengthen the primary agent's management layer to enforce per-subagent tool budgets, validate outputs against a defined schema before passing them forward, and establish explicit handoff contracts between stages.

D.  

Collapse the three subagents into a single large-context model call that handles retrieval, analysis, and generation in one pass.

A.  

Use prompt caching to reuse the static portions of the conversation context across turns instead of paying for them at the normal input-token rate on every request the application sends.

B.  

Switch to the Batch API for every turn so the per-call cost is reduced, treating the batch as the team's primary cost-control mechanism for multi-turn work.

C.  

Truncate every conversation to the last two messages so that token usage stays low and costs remain predictable across the application's normal operation.

D.  

Summarize each conversation after every turn and replace the full history with the summary on the next request, reducing token count at the cost of fidelity.

A.  

Add more tools to cover every variation of the requests the agent handles, on the grounds that more tools give the agent more accurate options to choose from.

B.  

Remove all tools and rely on the agent's general capability instead, with the application losing the workflows that previously relied on tools.

C.  

Add detailed examples to each tool's description so the agent can match incoming requests to the right tool by example, treating the examples as the team's selection mechanism.

D.  

Restructure the tool set by consolidating overlapping tools, removing unused tools, and clarifying tool descriptions so each tool has a distinct purpose.

A.  

A shared spreadsheet that lists configuration values for team members to reference and update by hand as the project evolves.

B.  

Environment variables that each developer sets on their own machine when working with Claude Code on the project.

C.  

A system prompt embedded in every Claude Code conversation by each developer at the start of every session in the project.

D.  

The settings.json file, scoped at the project level so the configuration applies consistently across developers and persists with the repository.

A.  

The third-party abstraction layer, on the grounds that multi-provider support is valuable for any application as a matter of long-term flexibility across vendors.

B.  

Both integration paths in parallel, where the application uses each path on different runs to compare which performs better in production over time.

C.  

A custom multi-provider abstraction layer the team builds in-house so that the team controls every part of the abstraction the application uses for its API calls.

D.  

Direct integration with Claude through its SDK, because no multi-provider need exists and abstraction would add complexity that does not pay off.

A.  

Add more explicit instructions so the most recent rules dominate the model's interpretation of the prompt during each request.

B.  

Audit the prompt for overlap, contradiction, and obsolete content, then refactor so each instruction is clear, current, and non-redundant.

C.  

Tighten only the contradictory rules first, treating the overlap and obsolete content as lower-priority work the team can address later.

D.  

Split the prompt across multiple system prompts so the model sees a smaller portion at any given time.

A.  

Retiring the system, since shipping is the final lifecycle stage for any application that reaches production.

B.  

Running a formal post-deployment review that assesses the development phase before the team does any further work on the system.

C.  

Operating and maintaining the system, including monitoring, responding to issues, and planning evolution.

D.  

Handing the system over to a separate operations team that will manage it independently of the development team going forward.

A.  

Move all instructions to a separate documentation site that developers consult during Claude Code sessions across all repositories.

B.  

Use a CLAUDE.md hierarchy that scopes general standards broadly and project-specific context within each repository's local CLAUDE.md.

C.  

Stop using CLAUDE.md altogether and ask each developer to configure Claude Code manually for each project they work on.

D.  

Use a single repository's CLAUDE.md as the central source of truth and link to it from every other repository's CLAUDE.md file.

A.  

Lower the model's temperature and re-run the inconsistent inputs to determine whether the parameter change reduces output variability across runs.

B.  

Retry every request that produces an unexpected output and log the results to identify whether repeated calls converge on a consistent response pattern.

C.  

Assume inconsistent outputs reflect normal LLM non-determinism and document the variability as an accepted characteristic of the application's behavior.

D.  

Capture full traces of input, system prompt, user messages, model output, and parameters, then analyze the differences between consistent and inconsistent runs.

A.  

Put the keys in the application's configuration file and check the configuration file into the team's repository alongside the rest of the source code.

B.  

Store the keys in a secrets manager or environment-specific configuration that is not checked into source code, and load them at runtime.

C.  

Use a single shared key across all external services, so any developer working on the application can find the keys easily during development.

D.  

Email the keys to each developer as needed and have each developer paste the keys into their local environment when they begin working on the application's code.

A.  

Examine production traces to identify whether the issue is hallucination by the model, context loss, prompt injection, or another failure mode before recommending a fix.

B.  

Replace the current model with a larger one to reduce the chance of hallucination, on the grounds that larger models tend to hallucinate less in typical applications.

C.  

Apply a retrieval-augmented generation pattern to ground the responses in source content before any further investigation of the production traces.

D.  

Add a system prompt instruction telling the model not to invent information, on the grounds that prompt-level instructions are the fastest fix for hallucination concerns.

A.  

Use multi-shot prompting by including the labeled examples in the prompt so Claude can match the desired output format on each request.

B.  

Use zero-shot prompting and rely on the model's general capability to produce the desired output format consistently across all incoming requests.

C.  

Embed the examples in a database and retrieve them at runtime as reference material for the team.

D.  

Train a custom model on the labeled examples before deployment so that the application does not need to include the examples in any prompt during operation.

A.  

As a custom tool embedded in this team's agent only, with other teams able to copy the implementation into their own agents when they need the capability.

B.  

As a shared library that each team imports into its own Claude application code, with each team responsible for keeping the library up to date in its integration.

C.  

As a Skill or MCP server because both are purpose-built for cross-team reuse independently by each consuming team.

D.  

As a wrapper around an existing built-in tool that adds the missing functionality, on the grounds that built-in tools cover the reuse pattern when extended carefully.

A.  

Run the agent only during business hours when humans are available to monitor its activity, treating the schedule as the primary control mechanism for destructive operations.

B.  

Add Claude hooks that intercept database operations and apply deterministic checks, such as blocking destructive queries or requiring approval, before the queries execute.

C.  

Switch to a higher-capability Claude model on the grounds that a more capable model is less likely to run destructive queries during normal operation across all requests.

D.  

Add a system prompt instruction telling the agent to be careful with database operations on every request the application handles during normal operation across all incoming traffic.

A.  

Add a system prompt instruction telling the model to produce schema-conforming arguments, treating the prompt instruction as the primary mechanism for keeping the agent's tool calls valid.

B.  

Keep the schema strict, validate arguments before dispatching, and return a structured error so the agent can retry.

C.  

Remove the schema entirely and rely on the downstream service to reject invalid calls, treating the downstream service as the team's primary enforcement layer.

D.  

Loosen the schema as the teammate proposed so the downstream service receives every call the agent makes during normal operation.

A.  

Validated and sanitized before being incorporated into prompts, and treated as data the model should not interpret as instructions.

B.  

Treated identically to trusted input from internal sources, on the grounds that all input the application receives can be handled the same way.

C.  

Routed through a separate Claude application with more lax security controls so the original application does not handle the untrusted input directly.

D.  

Blocked entirely so the application accepts only trusted input, with no path for legitimate external input the application might process.

A.  

Move all instructions into the user message so the application's behavior is controlled at the input point.

B.  

Consolidate behavioral and format instructions in the system prompt and keep user messages focused on the user's input.

C.  

Place each instruction wherever the developer who first wrote it found most natural, retaining the original intent.

D.  

Repeat all instructions in both the system prompt and the user message so the model sees each instruction more than once.

A.  

Define explicit session expiration rules, state reset triggers, and rules for starting fresh sessions so stale context does not leak into new conversations.

B.  

Define a single short session timeout that applies across all conversations and treat the timeout as the application's complete session lifecycle mechanism.

C.  

Define explicit session expiration rules but rely on users to start new conversations when they want fresh context, with no automatic reset triggers in the application.

D.  

Define state reset triggers tied to specific application events but apply them across all sessions globally, with no per-session expiration rules.

A.  

Approve the pull request and add the missing pieces yourself in a follow-up commit so the teammate can move on to other work immediately.

B.  

Approve the pull request as-is, on the grounds that the feature works in the happy path and the missing pieces can be added in follow-up commits.

C.  

Request changes that move prompt and model configuration to a configurable location and add tests, treating the missing error handling as a follow-up release item.

D.  

Request changes that move prompt and model configuration to a configurable location, add error handling for Claude API failures, and add tests for the integration.

A.  

Suggest replacing the descriptions with example calls embedded in the tool definition, treating example calls as a complete substitute for the prose description.

B.  

Agree with the teammate because tool names are sufficient for the model to choose the right tool on every request the agent handles.

C.  

Explain that the model uses the tool description to decide when to call the tool, and descriptions disambiguate cases where the tool name is not enough.

D.  

Suggest moving the descriptions out of the tool definition and into a separate documentation file the team maintains so the tool definitions stay short and the descriptions remain available.

A.  

Ask the product team to specify quality, latency, and cost targets, then select the model whose tradeoffs best fit those targets.

B.  

Run every Claude model on a representative sample and pick whichever scores best on a generic benchmark.

C.  

Choose a mid-tier model and ship the feature, because mid-tier models work for most use cases without specified targets.

D.  

Choose the largest, highest-capability Claude model, on the grounds that "best results" is most likely to mean highest quality.