Read the project's go.mod. Use only APIs available in its Cobra version; do not upgrade a dependency just to match this skill. Write Go for the version in its go directive (if a modern-Go guidelines skill such as use-modern-go is installed, follow it for that version).
Keep an existing working CLI. Migrate to or from Cobra only when asked or when the task needs it.
Build or change the command tree following Implementation choices. For a new command, start from the constructor and test pattern in references/commands.md.
Check the change against Flags and lifecycle pitfalls.
Before finishing, run the checks in Verify behavior.

Implementation choices

A new command starts from a constructor that takes its services and returns a fresh tree:

// examples/cobra/cli/cli.go (trimmed: text output, empty results, and completion omitted)
func NewRoot(list List, tags Tags) *cobra.Command {
	root := &cobra.Command{
		Use:           "catalog",
		SilenceErrors: true, // the process boundary owns error reporting
		SilenceUsage:  true, // explicit --help still works
	}
	var tag string
	ls := &cobra.Command{
		Use:  "list",
		Args: cobra.NoArgs,
		RunE: func(cmd *cobra.Command, _ []string) error {
			names, err := list(cmd.Context(), tag)
			if err != nil {
				return err
			}
			return json.NewEncoder(cmd.OutOrStdout()).Encode(names)
		},
	}
	ls.Flags().StringVar(&tag, "tag", "", "Filter by tag")
	root.AddCommand(ls)
	return root
}

A test builds its own tree, sets SetArgs, SetOut, and SetErr, runs ExecuteContext(t.Context()), and asserts on the captured buffers. The full version, with text output, completion, and tests, is in references/commands.md.

Construct fresh command trees with NewRoot(...) *cobra.Command (or an error-returning constructor if setup can fail). Inject services and writers; avoid package-level commands, mutable options, and registration in init(). Fresh instances prevent flags and tests from leaking state between executions.
Keep handlers thin. A CLI and a TUI over the same program should call the same core operations.
Put operations in RunE, with explicit Args validation (NoArgs, ExactArgs, MinimumNArgs, etc.). Without Args, a leaf command accepts arbitrary arguments, and a nested command group that has no RunE prints help for app group typo and exits 0. A script cannot tell that from success.
Use Flags() for a command's own options and PersistentFlags() for options genuinely shared by descendants. Parent local flags are not inherited.
Use ExecuteContext(ctx) and pass cmd.Context() into core operations. Define signal handling and exit codes at the process boundary. Cobra does not choose your cancellation or partial-failure exit codes.
Write payloads through cmd.OutOrStdout() and diagnostics through cmd.ErrOrStderr(). Keep JSON output free of progress, styling, and prose. Do not use cmd.Print* for payloads: it writes to the SetOut writer when one is set and otherwise to the process's stderr, so a test that calls SetOut passes while the real binary sends the payload to stderr.
Assign error printing to one layer. If the caller prints errors, set SilenceErrors: true; use SilenceUsage deliberately so runtime failures do not emit a help page. Still return the error. Do not os.Exit, log.Fatal, or cobra.CheckErr inside reusable commands.
Create dependencies only when the operation needs them. --help, --version, and completion should not require a valid config, network login, or a writable state database merely because the command tree was constructed.

Flags and lifecycle pitfalls

Required flags and flag groups concern flags explicitly supplied by the user (Changed), not whether an effective value is nonzero. If config or env can satisfy a required setting, validate the resolved setting instead.
Distinguish --enabled=false from an omitted flag using Changed. Prefer StringArray for repeated literal values, StringSlice for CSV/repeated values, and Count for -v, -vv, etc.
Check errors from flag marking and completion registration. Group helpers such as MarkFlagsMutuallyExclusive can panic when given an undefined flag name; register flags first.
Persistent hooks are nearest ancestor wins by default, including hooks on the command itself. A child's persistent hook can bypass root configuration loading. Prefer explicit shared setup over surprising hook inheritance.
A root PersistentPreRunE is skipped for --help and --version, but it runs for completion <shell>, every __complete request, and help <command>. If it loads config and fails, tab completion breaks wherever the config file is missing or invalid. Load config inside the operation that needs it, or return early from the hook for those commands.
cobra.EnableTraverseRunHooks is a process-global opt-in: persistent pre-hooks then run root-to-leaf and post-hooks leaf-to-root. Do not toggle it per command or in parallel tests. TraverseChildren is a different flag-parsing option.
Args validation runs first, then pre-hooks, then required-flag and flag-group validation. Keep mutations in RunE. Post-hooks do not run after a failed RunE; acquire resources and defer cleanup in the same operation instead.
When combining Viper, bind flags after defining them, then read the effective values from Viper after parsing. Viper does not update variables bound with StringVar from config or environment values.

Task-specific references

references/commands.md: the default starting pattern. A constructor with injected services, context, output writers, the process boundary, and command tests.
references/completion.md: shell completion, generated docs, and a checklist for debugging routing and flags.
For file/env/flag precedence and decoding, use the viper skill if it is installed. Without it, consult the upstream Viper API; do not add Viper merely because a command uses Cobra.

references/commands.md also covers Version, help groups, hidden and deprecated commands, and flag groups. For MatchAll, custom pflag.Value types, and the rest of the API, see the API documentation under Sources.

Verify behavior

Before finishing, test with fresh command instances using SetArgs, SetIn, SetOut, SetErr, and ExecuteContext, and check:

invalid input does not invoke the service;
output lands on the right stream;
cancellation reaches the service;
explicit false and zero values work;
ancestor options and config behave on leaf commands;
custom completion returns the expected candidates and directive.

Avoid snapshots of entire generated help pages unless their exact layout is a requirement. Run go vet and go test for the packages you changed. When a check fails, find the matching entry in Flags and lifecycle pitfalls, fix the command, and rerun every check.

Sources

Verified against cobra v1.10.2 and pflag v1.0.9; each stated behavior has a test in the examples module. When the project's version differs, trust go doc github.com/spf13/cobra.Command and its module source over this skill.

cobra

Cobra

Workflow