Anthropic CCA-F Questions & Answers

The normal policy remains valid for other files. Ordinary files still need the baseline convention even if generated files use a documented exception. A general naming policy and a generated-code exception both match the same file. Instead, identify the overlapping glob scopes and specify an explicit narrow exception to the general rule. Identifying overlapping globs and documenting the exception prevents contradictory naming demands on the same generated artifact.

Other developers would still see the overlap. Teammates lacking that private override will still receive the conflicting common rules. A general naming policy and a generated-code exception both match the same file. Instead, identify the overlapping glob scopes and specify an explicit narrow exception to the general rule. Identifying overlapping globs and documenting the exception prevents contradictory naming demands on the same generated artifact.

Conflicting instructions leave generated output ambiguous. Two active incompatible rules will send contradictory instructions to the model for one artifact. Two matching rules can be simultaneously relevant, so their instructions must state how the narrower generated-code exception interacts with the general rule. A general naming policy and a generated-code exception both match the same file. Instead, identify the overlapping glob scopes and specify an explicit narrow exception to the general rule.

This defeats scope rather than covering the added test tree. A global wildcard triggers the test policy in production and documentation changes too. Changing folder structure to satisfy an obsolete glob hides the configuration defect and disrupts otherwise valid project layout. A repository adds a second test directory outside the current glob. Instead, expand the test-file pattern to cover the new location without activating unrelated files.

Restructuring working files for instruction matching is unnecessary. Renaming source folders can break imports or builds while hiding the true configuration problem. Changing folder structure to satisfy an obsolete glob hides the configuration defect and disrupts otherwise valid project layout.

The new test files still fail conditional loading. README advice is informational; it does not cause the loader to match the added directory. A repository adds a second test directory outside the current glob. Instead, expand the test-file pattern to cover the new location without activating unrelated files.

There is no shared migration sequence or rollback. Independent patches may be deployed in an order that breaks shared dependencies and rollback. A cross-service library upgrade has dependency ordering and compatibility constraints that must be mapped before code changes. A library upgrade will affect dozens of services and involves incompatible interfaces.

Dependency contracts may differ across services. Services can have different request and response contracts, making one guessed type unsafe. A library upgrade will affect dozens of services and involves incompatible interfaces. Instead, explore design and dependencies in plan mode before implementation.

Cross-service breakages become expensive after mutation. Changes across consumers become expensive to reverse when incompatible interfaces are discovered late. A library upgrade will affect dozens of services and involves incompatible interfaces. Instead, explore design and dependencies in plan mode before implementation.

The defect is localized to a single function. Global edits multiply the affected code surface without evidence of defects in those other functions. A single input validator compares the wrong constant and a regression test is ready. Instead, apply the focused fix without unnecessary architectural planning.

The known validator issue needs limited execution. A broad migration review cannot materially improve a known one-line correction already covered by a test. A single input validator compares the wrong constant and a regression test is ready. Instead, apply the focused fix without unnecessary architectural planning.

Behavioral verification is still required after direct execution. A plausible change may still invert a comparison or mishandle boundary input unless executed tests pass. A known one-line input-validator defect with an existing regression test is a bounded execution task, not an unresolved architecture decision. A single input validator compares the wrong constant and a regression test is ready.

Planning must remain separate from authorized mutation. The plan may describe DDL but must not become authorization to execute it on customer records. A schema change touches customer records, backups, and multiple consumers. Instead, map migration strategy and rollback before authorized execution.

This risks data loss and broken consumers. Early deletion can strand application readers and destroy data before migration compatibility is checked. A schema change touches customer records, backups, and multiple consumers. Instead, map migration strategy and rollback before authorized execution.

Recovery time and dependent services still matter. Backups alone do not prove acceptable downtime, restore integrity or consumer compatibility. Plan mode can reveal retention and privacy implications while the actual destructive change remains subject to authorization. A schema change touches customer records, backups, and multiple consumers.

Unreviewed changes waste resources and increase risk. Running both designs in production creates cost and risk before selecting acceptance criteria. Implementing both in production introduces unnecessary risk and does not replace an agreed decision record. Two viable architectures have different deployment and observability tradeoffs. Instead, use plan mode to compare alternatives and agree requirements.

Unstructured preference ignores requirements and tradeoffs. The model's first suggestion is not a measured comparison against requirements. Two viable architectures have different deployment and observability tradeoffs. Instead, use plan mode to compare alternatives and agree requirements. Implementing both in production introduces unnecessary risk and does not replace an agreed decision record.

One dimension cannot settle competing architectural requirements. A smaller footprint may sacrifice mandatory tracing or deployment resilience required by the owner. Two viable architectures have different deployment and observability tradeoffs. Instead, use plan mode to compare alternatives and agree requirements. Implementing both in production introduces unnecessary risk and does not replace an agreed decision record.

The typo is isolated and has no inter-team contract impact. Involving unrelated owners creates coordination delay without a dependency or approval requirement. A migration plan or function rewrite would add scope and regression risk to a trivial text modification. Instead, make the direct text correction.

Unrelated code modifications amplify regression risk. Unrequested code edits introduce test failures into code that was previously stable. A migration plan or function rewrite would add scope and regression risk to a trivial text modification. A docstring contains one misspelling and no behavior or tests will change. Instead, make the direct text correction.

The edit has no architectural uncertainty. An architecture migration plan adds paperwork when no consumer behavior or interface is changing. A docstring contains one misspelling and no behavior or tests will change. Instead, make the direct text correction. A migration plan or function rewrite would add scope and regression risk to a trivial text modification.

Without paths the lead cannot validate dependencies. A pathless dependency summary cannot be used to open the module or verify the claimed import edge. Keeping raw search logs in the lead agent's context obscures the actual paths and risks needed for implementation. A large repository reconnaissance generates verbose tool logs that disrupt implementation reasoning. Instead, delegate verbose read-only exploration and return compact module and dependency findings to the lead agent.

This reproduces the original context overload. Raw search logs recreate the exact prompt pollution that prevented the lead from reasoning. A restricted exploration agent can map dependency edges with read-only tools and return a compact, cited module inventory. A large repository reconnaissance generates verbose tool logs that disrupt implementation reasoning.

Reconnaissance should not mutate the codebase. Exploratory permissions should be read-only; mutation could disturb evidence before review. Keeping raw search logs in the lead agent's context obscures the actual paths and risks needed for implementation. Instead, delegate verbose read-only exploration and return compact module and dependency findings to the lead agent.

The policy requires deliberate review and approval. Security policy ownership cannot be delegated invisibly to model-generated code. A new authentication workflow requires design review but then contains clear implementation tasks. Instead, plan the authorization model first and execute approved substeps afterward.

Repeated planning delays safe implementation. Once design constraints are fixed, further planning does not make the accepted tasks safer. A new authentication workflow requires design review but then contains clear implementation tasks. Instead, plan the authorization model first and execute approved substeps afterward.

Premature code may enforce the wrong access model. Implementers could bake in a trust model that authorization stakeholders never approved. A new authentication workflow requires design review but then contains clear implementation tasks. Instead, plan the authorization model first and execute approved substeps afterward.

An unlocalized failure may be environmental rather than functional. Increasing retries changes symptoms without locating the race or external dependency. A flaky test fails only in one environment and the cause has not been localized. Instead, use exploratory plan/read-only diagnosis before choosing a code modification. An environment-specific flaky test needs diagnosis of nondeterministic inputs, timing and external dependencies before a patch is chosen.

The failure is tied to an unexamined environment. Local passing behavior does not resolve a defect observed only in a specific CI environment. An environment-specific flaky test needs diagnosis of nondeterministic inputs, timing and external dependencies before a patch is chosen.

Large edits obscure the actual failure mechanism. Rewriting all tests may delete the one failure needed to understand the environment. A flaky test fails only in one environment and the cause has not been localized. Instead, use exploratory plan/read-only diagnosis before choosing a code modification.

Broad distribution worsens privacy and governance risks. Unbounded replication makes it harder to enforce permitted storage locations and deletion requirements. A new region for user data has several deployment choices with legal constraints. Instead, assess tradeoffs in plan mode before changing infrastructure.

Suppressing records does not make placement lawful. Suppressing records removes evidence needed for audit but leaves the actual data movement unchanged. A new region for user data has several deployment choices with legal constraints. Instead, assess tradeoffs in plan mode before changing infrastructure.

A cost-only decision ignores residency restrictions. Financial savings cannot establish that a cloud region satisfies residence and transfer obligations. A new region for user data has several deployment choices with legal constraints. Instead, assess tradeoffs in plan mode before changing infrastructure.