chore: make AI instructions stronger

Signed-off-by: Ilya Lesikov <ilya@lesikov.com>

Author: ilya-lesikov

AGENTS.md

@@ -1,8 +1,12 @@
 # Repository Guidelines
 
+All rules in this document are requirements — not suggestions. ALWAYS follow them.
+
 Nelm is a Go-based Kubernetes deployment tool, which deploys Helm charts, is compatible with Helm releases, and is an alternative to Helm. Nelm is built on top of a Helm fork ([werf/3p-helm](https://github.com/werf/3p-helm)) and is also used as the deployment engine of [werf](https://github.com/werf/werf).
 
-## Commands
+## Commands (MANDATORY)
+
+ALWAYS use these `task` commands. NEVER use raw `go build`, `go test`, `go fmt`, or `golangci-lint` directly.
 
 - `task build` — Build binary for current OS/arch to `./bin/`. Accepts `pkg=...` to build a specific package.
 - `task format` — Run all formatters. Accepts `paths="./pkg/..."` to scope to a specific package.
@@ -33,28 +37,28 @@ Nelm is a Go-based Kubernetes deployment tool, which deploys Helm charts, is com
 - `internal/ts/` — TypeScript support (bundling, rendering).
 - `internal/util/` — internal utility functions.
 
-## Testing
+## Testing (MANDATORY)
 
-- When writing new tests → use `testify` (`assert`, `require`, `suite`).
-- When writing tests as an AI agent → name the file `*_ai_test.go`, add `//go:build ai_tests` build tag, prefix test functions with `TestAI_`.
-- Place tests alongside source files, not in a separate directory.
+- ALWAYS use `testify` (`assert`, `require`) when writing new tests.
+- When writing tests as an AI agent → ALWAYS name the file `*_ai_test.go`, add `//go:build ai_tests` build tag, prefix test functions with `TestAI_`.
+- ALWAYS place tests alongside source files, not in a separate directory.
 - Test helpers go in `helpers_test.go` (or `helpers_ai_test.go` for AI-written helpers).
 - Test fixtures go in `testdata/` subdirectory next to the tests.
 - Shared test helpers are in `internal/test/`.
 
-## Work standards
+## Work standards (MANDATORY)
 
-- Always use `task` commands for build/test/lint/format — never raw `go build`, `go test`, `go fmt`, or `golangci-lint` directly.
-- When logging → use `log.Default` from `pkg/log`. Never use `fmt.Println`, `slog`, or `logrus` directly.
-- Read and strictly follow the project code style defined in [CODESTYLE.md](CODESTYLE.md).
-- Verify, don't assume — always check the actual state before making changes.
-- Don't leave TODOs, stubs, or partial implementations.
+- ALWAYS use `task` commands for build/test/lint/format — NEVER raw `go build`, `go test`, `go fmt`, or `golangci-lint` directly.
+- ALWAYS use `log.Default` from `pkg/log` for logging. NEVER use `fmt.Println`, `slog`, or `logrus` directly.
+- ALWAYS read and strictly follow the project code style defined in [CODESTYLE.md](CODESTYLE.md). Every rule in it is a requirement.
+- ALWAYS verify, don't assume — check the actual state before making changes.
+- NEVER leave TODOs, stubs, or partial implementations.
 
-## PR review guidelines
+## PR review guidelines (MANDATORY)
 
-- Do not add new external dependencies without flagging to the user first.
-- Do not introduce breaking user-facing changes (not API changes) unless they are hidden behind a feature flag. Flag to the user first.
-- Do not introduce changes that may compromise security. Flag to the user first.
+- NEVER add new external dependencies without flagging to the user first.
+- NEVER introduce breaking user-facing changes (not API changes) unless they are hidden behind a feature flag. Flag to the user first.
+- NEVER introduce changes that may compromise security. Flag to the user first.
 
 ## Related repositories
 

CLAUDE.md

@@ -1,25 +1,31 @@
 # Claude Code Instructions
 
-Follow @AGENTS.md for universal project guidelines.
+ALWAYS strictly follow @AGENTS.md for universal project guidelines.
 
-## Exploring the codebase
+## Exploring the codebase (MANDATORY)
 
-- To find definitions, references, callers, or implementations → use **LSP** (`goToDefinition`, `findReferences`, `incomingCalls`/`outgoingCalls`, `workspaceSymbol`).
-- To find code by intent or understand how something works → use **CodeAlive MCP** (`codebase_search`, `codebase_consultant`). Call `get_data_sources` first.
-- To explore broad architecture across many files → use **Task/Explore** subagent as a last resort.
+These are NOT suggestions — they are requirements. NEVER skip them in favor of Grep/Glob.
 
-## External documentation
+- ALWAYS use **LSP** (`goToDefinition`, `findReferences`, `incomingCalls`/`outgoingCalls`, `workspaceSymbol`) to find definitions, references, callers, or implementations. NEVER use Grep for these — LSP is precise, Grep is a guess.
+- ALWAYS use **CodeAlive MCP** (`codebase_search`, `codebase_consultant`) to find code by intent or understand how something works. Call `get_data_sources` first. NEVER substitute this with Grep-based keyword searching when the query is semantic/intent-based.
+- Only fall back to **Grep/Glob** for simple literal pattern matching (e.g., finding a specific string, config key, or error message).
+- Only use **Task/Explore** subagent as a last resort when the above tools are insufficient.
 
-- To look up documentation for any technology, library, or tool → use **Context7 MCP** (`resolve-library-id` + `query-docs`).
+## External documentation (MANDATORY)
 
-## E2E testing
+These are NOT suggestions — they are requirements. NEVER skip them in favor of Grep.
 
-After making changes, verify them end-to-end against the local dev cluster:
+- ALWAYS use **LSP** `hover` to look up documentation, type signatures, and godoc for any symbol in the codebase (including symbols from external dependencies).
+- ALWAYS use **Context7 MCP** (`resolve-library-id` + `query-docs`) to look up broader documentation, guides, or examples for any technology, library, or tool. NEVER guess at APIs — look them up.
+
+## Verifying changes (MANDATORY)
+
+After making changes, ALWAYS verify them end-to-end against the local dev cluster:
 
 1. `task build` — build the binary.
 2. Run the built binary from `./bin/nelm` against the cluster to deploy/test.
-3. Use **Kubernetes MCP tools** (`mcp__kubernetes__*`) to inspect the cluster state — get resources, describe them, check logs, exec into pods, etc. Prefer these over raw `kubectl` via Bash.
+3. ALWAYS use **Kubernetes MCP tools** (`mcp__kubernetes__*`) to inspect the cluster state — get resources, describe them, check logs, exec into pods, etc. NEVER use raw `kubectl` via Bash when an MCP tool can do it.
 
-## Code style
+## Code style (MANDATORY)
 
-- When writing or modifying Go code, strictly follow @CODESTYLE.md.
+- ALWAYS strictly follow @CODESTYLE.md when writing or modifying Go code. Every rule in it is a requirement. Read it before writing any code if you haven't already in this session.

CODESTYLE.md

@@ -1,3 +1,5 @@
+> **AI agents**: all rules below are mandatory. NEVER deviate from them.
+
 ## Design
 
 * Prefer stupid and simple over abstract and extendable.