Agentic Codebase Skill

Overview

Assess and improve a codebase's readiness for AI agent collaboration. Based on the 15-principle Agentic Codebase Manifesto organized across 3 pillars.

This skill operates in two modes:

Mode: audit -- Score the codebase against all 15 principles, produce a report with evidence and prioritized actions
Mode: setup -- Run audit first, then incrementally fill gaps with real content generated from codebase analysis

Mode is set by the calling command (`/compound:agentic-audit` or `/compound:agentic-setup`). The command wrapper tells you which mode to run -- do not parse `$ARGUMENTS` for mode detection.

Stack Detection

Before auditing, detect the project stack to adapt checks:

Look for package.json (Node/TS), pyproject.toml or setup.py (Python), Cargo.toml (Rust), go.mod (Go), Makefile, CMakeLists.txt (C/C++)
Check for framework markers: next.config, django, fastapi, express, etc.
Identify build/test/lint commands from config files
Store detected stack for use in principle checks and AGENTS.md generation

Audit Methodology

Scoring Rubric

Each principle is scored:

0 (Absent): No evidence of this principle in the codebase
1 (Partial): Some evidence but incomplete or inconsistent
2 (Present): Clear, consistent implementation

Adapt criteria to the detected stack. For example: "strict mode" means TypeScript strict, Python mypy --strict, or Rust default safety. "Linter" means ESLint, pylint/ruff, clippy, golangci-lint, etc. Score based on the ecosystem's equivalent tooling.

The 15 Principles

Pillar I: Codebase Memory (Traceability) -- max 8 points

P1. Repository is the only truth Check: All context an agent needs lives in version control Evidence: Look for docs/ directory, inline documentation, config files Score 0: No docs directory, no README beyond boilerplate Score 1: README exists but key context lives elsewhere Score 2: Comprehensive docs/, config, and context all in-repo

P2. Trace decisions, not just outcomes Check: Architectural decisions have recorded rationale Evidence: Look for docs/adr/, docs/decisions/, ADR files Score 0: No decision records Score 1: Some decisions documented but inconsistent format Score 2: ADR directory with structured records

P3. Never answer the same question twice Check: Solutions/fixes are documented to prevent rediscovery Evidence: Solutions docs, post-mortems, troubleshooting guides, or a memory system Score 0: No solutions documentation Score 1: Scattered notes but no systematic approach Score 2: Structured solutions docs or integrated memory system

P4. Knowledge is infrastructure Check: Documentation is versioned alongside code Evidence: Specs, research, standards co-located in repo Score 0: Documentation lives outside version control Score 1: Some docs in repo but key knowledge is external Score 2: All project knowledge versioned in docs/

Pillar II: Implementation Feedbacks (Mechanical Verification) -- max 8 points

P5. Test is specification Check: Tests define behavior before or alongside implementation Evidence: Test files, coverage tooling, test-first patterns Score 0: No tests or minimal coverage Score 1: Tests exist but post-hoc or inconsistent Score 2: Comprehensive test suite with test-driven patterns

P6. Constraints are multipliers Check: Linters, type checkers, architectural rules configured and enforced Evidence: ESLint/pylint/clippy config, TypeScript strict mode, CI enforcement Score 0: No linting or type checking Score 1: Linter exists but not enforced in CI Score 2: Strict linting + type checking enforced in CI

P7. Write feedback for machines Check: Error messages, logs, and output are structured for agent consumption Evidence: Structured logging, clear error messages with context Score 0: Unstructured logs, generic error messages Score 1: Some structured logging but inconsistent Score 2: Structured logging throughout, remediation hints in errors

P8. Fight entropy continuously Check: Active maintenance processes prevent drift Evidence: Automated formatting, dependency updates, quality monitoring Score 0: No automated maintenance Score 1: Basic formatting but no proactive monitoring Score 2: Automated formatting + dependency updates + quality tracking

Pillar III: Mapping the Context (Navigable Structure) -- max 8 points

P9. Map, not manual Check: Entry point document provides a navigable map, not an encyclopedia Evidence: AGENTS.md, CLAUDE.md, or similar Score 0: No agent-facing entry point document Score 1: README exists but not optimized for agents Score 2: Dedicated AGENTS.md or CLAUDE.md with commands, structure, conventions

P10. Explicit over implicit, always Check: Types, naming, patterns are explicit Evidence: Type annotations, consistent naming, documented conventions Score 0: No type annotations, inconsistent naming Score 1: Some types but gaps, or undocumented conventions Score 2: Full type coverage, documented naming conventions

P11. Modularity is non-negotiable Check: Single responsibility per file, clear boundaries Evidence: File sizes, module organization, dependency structure Score 0: Monolithic files (>500 LOC common), unclear boundaries Score 1: Some modular structure but large files remain Score 2: Consistent small files, clear APIs, enforced boundaries

P12. Structure in layers, govern by inheritance Check: Layered architecture with explicit dependency rules Evidence: Layer separation, import rules, dependency graph Score 0: No discernible layering Score 1: Informal layers but no enforcement Score 2: Explicit layers with enforced dependency directions

Cross-Cutting -- max 6 points

P13. Simplicity compounds Check: Prefer boring technologies, minimal abstractions Evidence: Dependency count, abstraction depth Score 0: Over-engineered with many abstractions Score 1: Moderate complexity, some unnecessary abstractions Score 2: Minimal dependencies, straightforward patterns

P14. Human designs the system, not the output Check: Human effort in system design (tests, docs, constraints) Evidence: Quality of test harnesses, documentation, CI/CD Score 0: No investment in development infrastructure Score 1: Some tooling but gaps in key areas Score 2: Strong CI, testing framework, documentation system

P15. Parallelize by decomposition Check: Work can be split into independent units Evidence: Module independence, clear interfaces, minimal coupling Score 0: Tightly coupled, hard to work on independently Score 1: Some independent modules but shared state Score 2: Well-decomposed with clear interfaces

Audit Execution Steps

Run `ca search "agentic codebase"` for relevant lessons
Detect project stack (see Stack Detection above)
Use Glob and Grep to check for evidence of each principle:
- Glob for: docs/**, .test., test.go, test.py, _test.rs, .eslintrc, AGENTS.md, CLAUDE.md
- Grep for: type annotations, structured logging, ADR format
- Read key files: README, config files, sample source files
Score each principle (0-2) with specific evidence
Aggregate scores by pillar and compute total out of 30
Generate prioritized actions (score-0 first, then score-1)
Present report to user

Report Format

Present as markdown tables per pillar:

Pillar I: Codebase Memory -- X/8

#	Principle	Score	Evidence
P1	Repository is the only truth	0/1/2	finding
...repeat for all pillars with separator rows...

Overall Score: X/30

Priority Actions

[Score-0 items first, most impactful]
...

After presenting, use `AskUserQuestion`: "Create a beads epic with issues for improvements?" If yes, create epic via bd create and individual issues.

Setup Methodology

Prerequisites

Run the full audit first. Setup only addresses gaps found by the audit.

Setup Execution Steps

Present audit findings summary
For each principle scored 0 or 1, propose a concrete action:

P1/P4 gaps: Create docs/ skeleton (INDEX.md, adr/, standards/) with real content from analysis P2 gaps: Create ADR template and first ADR from actual architecture analysis P3 gaps: Suggest solutions documentation structure P5 gaps: Suggest test framework setup based on detected stack P6 gaps: Suggest linter/type checker configuration for detected stack P7 gaps: Suggest structured logging patterns for detected stack P9 gaps: Generate AGENTS.md by analyzing actual codebase (build commands, structure, conventions) P10 gaps: Suggest type annotation and strict mode settings P11 gaps: Identify files >500 LOC, suggest refactoring targets P12 gaps: Document layer structure and suggest import lint rules (e.g., eslint-plugin-import boundaries, Rust mod visibility) P13 gaps: Flag over-abstraction (deep inheritance, excessive wrappers), suggest simplification targets P14 gaps: Suggest CI pipeline improvements, test harness setup, or pre-commit hooks for detected stack P15 gaps: Identify tightly coupled modules, suggest interface extraction for parallel workability P8 gaps: Suggest automated formatting (prettier/black/rustfmt), dependency update tooling (renovate/dependabot), and quality monitoring

Before each action, use `AskUserQuestion`: "Create [file]? Preview: [content]"
Only create/modify files the user approves
Never overwrite existing files without explicit approval

Setup Completion Gate

After all approved actions are applied, verify:

List all files created/modified during setup
Run quality gates if available (`pnpm test`, `pnpm lint`)
Confirm no existing files were overwritten without approval
Present summary: principles addressed, files created, remaining gaps

Memory Integration

Before analysis: `ca search "agentic codebase"` for relevant lessons
After completing: offer `ca learn` to capture insights

Common Pitfalls

Scoring too generously without specific evidence for score 2
Generating template content instead of analyzing the actual codebase
Overwriting existing files without asking
Not detecting the project stack before generating content
Creating too many files at once instead of prioritizing
Forgetting to offer beads epic creation after audit

Quality Criteria

All 15 principles assessed with specific evidence
Scores justified with findings
Pillar totals and overall score calculated correctly
Actions prioritized (score-0 before score-1)
Stack detected and checks adapted accordingly
User consulted via AskUserQuestion at key decisions
Memory searched before analysis
Setup mode ran audit first
No files overwritten without approval
Generated content based on actual codebase analysis

Agentic Codebase

SKILL.md.claude/skills/compound/agentic/SKILL.md