In one line: Select tools by model capability and development phase, not by habit — this section is the decision matrix.
The methodology integrates plugins, MCP servers, and skills into a unified stack. Tool selection is capability-based.
Capability Assessment
Before selecting tools, assess the AI model's capabilities:
| Capability | How to Assess | Threshold |
|---|
| Context window | Check model documentation | Large (>200K), Medium (50-200K), Small (<50K) |
| Internal reasoning | Does the model have thinking/reasoning blocks? | Yes / No |
| Tool reliability | Can the model use tools (Read, Edit, Bash) without errors? | Reliable / Unreliable |
| Multi-step planning | Can the model decompose and execute 5+ step plans? | Yes / Needs guidance |
| Code understanding | Can the model read a 500-line file and modify precisely? | Full context / Needs navigation |
Code Navigation
| Capability | Recommendation |
|---|
| Reliable direct inspection within the task budget | Read scoped source and configuration; preserve authoritative references. |
| Cross-file navigation or dependency analysis | Evaluate approved symbol/graph tools such as Serena or codebase-memory-mcp; verify index freshness and coverage. |
| Context pressure or incomplete retrieval | Reduce/split scope or use a tested retrieval binding; do not infer that a named plugin is mandatory from context-window size. |
Reasoning Support
| Capability | Recommendation |
|---|
| Planning and difficult decisions | Evaluate observable task outcomes, counterexamples and concise decision rationale; visible reasoning blocks do not determine correctness. |
| Optional decomposition tools | Adopt only where they improve measured work within approved data/tool boundaries; they are not an audit prerequisite. |
| Audit trail required | Store concise rationales, evidence references, relevant outputs and approval events under the applicable retention policy; not hidden reasoning chains. |
Code Quality
| Layer | Tool | When |
|---|
| Real-time | TypeScript LSP, Pyright LSP | During implementation |
| Pre-commit | ruff (lint), pyright (types), tsc (TypeScript) | Every commit |
| Security | Aikido plugin (SAST, secrets) | During development |
| PR review | Approved code-review integration | Before merging, on the actual candidate |
| PR security | claude-code-security-review GitHub Action | On every PR |
| Post-merge | code-simplifier (bloat detection) | After major features |
These are reference-host examples, not universal vendor or cadence requirements. The approved stack/evidence profile selects actual tools, versions, subject scope, source-processing permissions and required execution. A package name or model label does not establish installation or completion.
Brainstorm Phase
| Tool | Purpose | Capability Requirement |
|---|
| Visual companion (browser) | Mockups, diagrams | Any |
| Terminal discussion | Conceptual choices | Any |
| Architecture index lookup | Understand existing code | Any |
Existing-system brainstorming often needs source and integration discovery. Use authorized read-only evidence where it answers the question; a new-product discussion may need none.
Design Phase
| Tool | Purpose | Capability Requirement |
|---|
| Architecture index | Find relevant doc pages | Any |
Serena get_symbols_overview | Understand existing interfaces | Medium/small context |
| Read tool | Read existing files directly | Large context |
Plan Phase
| Tool | Purpose | Capability Requirement |
|---|
Architecture index tests: field | Identify test files | Any |
Architecture index component_map | Map files to tasks | Any |
Implement Phase
| Tool | Purpose | Capability Requirement |
|---|
| Read / Edit / Bash | Direct file manipulation | Large context, reliable tools |
Serena find_symbol / replace_symbol_body | Precise symbol editing | Medium/small context |
| TypeScript LSP | Frontend type checking | Any (if available) |
| Context7 | Library documentation lookup | Any (prevents hallucinated APIs) |
| Temporal MCP | Workflow debugging | Projects using Temporal |
| Neo4j MCP | Graph queries | Projects using Neo4j |
Review Phase
| Tool | Purpose | Capability Requirement |
|---|
| code-review plugin | Scoped candidate review | Approved model/tool binding |
| Domain-specific review agents | Compliance, security, API, migration | Project-specific |
| CodeRabbit | External AI review perspective | Optional; approved source processing and current service terms |
Verify Phase
| Tool | Purpose | Capability Requirement |
|---|
| Stop hook | Debug-only checklist in the reference adapter | Verify host delivery separately |
Architecture index tests: | Declared mapping; compare against actual execution | A mapping alone is not a receipt |
| ruff / pyright / tsc | Lint + type check | Any |
| Aikido | Security scan | Any (if configured) |
All phase tables are illustrative capability mappings. The general architecture index is an adopter extension, not a supplied plugin; tests metadata is not an execution receipt. Context thresholds, pricing, package names and model selections must be verified for the actual binding rather than treated as fixed acceptance rules.
Plugin Stack Reference
| Plugin | Adoption reference | Purpose |
|---|
| Superpowers | Verify package, version and host procedure under §4.4 | Lifecycle skills |
| code-review | Select the approved review integration | Scoped agent-assisted review |
| code-simplifier | Optional, evaluated helper | Refactoring suggestions requiring verification |
Recommended (install based on stack)
| Plugin | Adoption reference | When |
|---|
| typescript-lsp | Verify approved host package | Projects selecting TypeScript language services |
| pyright-lsp | Verify approved host package | Projects selecting Python language services |
| Serena | Verify approved host package and index | Tasks benefiting from symbol navigation |
| Server | Package | When |
|---|
| Context7 | Verify approved provider/package | Optional library-documentation retrieval |
| Temporal MCP | temporal-mcp | Projects using Temporal |
| Neo4j MCP | @johnymontana/neo4j-mcp | Projects using Neo4j |
| codebase-memory-mcp | codebase-memory-mcp | Large codebases, token optimization |
| Sequential Thinking | sequential-thinking | Optional structured problem decomposition; not a prerequisite for an audit trail |
AGENTS.md
Use the selected host's supported instruction entry point and verify actual loading. Where two hosts support the same content, an approved pointer or symlink may avoid duplication; do not overwrite an existing AGENTS.md or assume every host follows links and precedence identically. Preserve tool-specific differences as explicit, bounded references.
Adoption Tier Impact
| Tier | Capability obligation | Binding choice |
|---|
| Core | Adopted authority, lifecycle, evidence and required controls | Existing approved tools or verified reference adapters; no universal vendor install list |
| Recommended | Navigation, language services and domain integration where useful | Evaluate against workload, permissions, privacy and maintenance cost |
| Extended | Additional automation/retrieval where justified | Expand only with scoped acceptance evidence and approved operating authority |
Competitive Landscape (March 2026)
The former comparison table and uniqueness claim did not establish a defined comparison corpus or independent outcomes. They are not a basis for adoption. S4U's intended value is a maintained combination of business-intent traceability, bounded engineering procedure and inspectable evidence. Evaluate that value locally and compare alternatives using the factory research; no market-wide superiority is claimed.