Imported from garyld1962/savviety-skills (kimi/skills/modernize/SKILL.md). Install upstream with npx skills add garyld1962/savviety-skills --skill modernize. Copyright stays with the author.

/skill:modernize — Code Modernization Audit (AI-Time Refresh)

Purpose: Read an existing codebase holistically and produce a prescriptive refactor plan calibrated to current AI toolchain capability. The output is a plan-shaped artifact that hands off to /skill:execute-prd --type=refactor so modernization actually happens — this skill is the front-end of a refactor pipeline, not a free-standing report generator.

When to Use

• Revisiting a project that has been dormant for a meaningful stretch (months+).
• A codebase that bears the marks of older-AI-era patterns: over-defensive try/catch, premature abstraction, type erosion (any/unknown overuse), kitchen-sink "utils" modules.
• After a major model or skill release where you want to capitalize on what's newly feasible (large multi-file consistent refactors, type-system surgery, cross-module data-flow tracing).
• Phrases: "modernize this", "refresh the code", "what would this look like today".

When NOT to Use

• Codebase is actively being developed; no AI-time gap exists.
• You have a specific bug or feature — use the relevant focused skill.
• You suspect the project wants a fundamentally different stack — that's a human decision; this skill stays inside the current platform.
• Single-file polish — make focused code edits directly.
• Dependency-driven modernization (CVEs, package upgrades) — use /skill:dep-audit + /skill:dep-migrate.

Scope Boundaries

Strictly within the current platform/stack.

• Will recommend: structural refactors, type improvements, module reorganization, error-handling sharpening, test depth, infra/config organization, dead-code removal, naming clarity.
• Will NOT recommend: changing the framework, rewriting the public API, switching languages, replacing the database, scope changes that touch >50 files, introducing major new dependencies.

If a "this whole thing wants a different framework" signal is overwhelming, the skill flags it as an out-of-scope-finding in the report's appendix. It does not include such items in the plan moves. Framework changes are human decisions.

Arguments

Workflow

Phase 1: Project shape detection

Per the _internal/modernization-rubric contract §1, classify:

• Language(s) — primary language by LOC if multiple manifests exist
• Type — CLI / library / app / service / monorepo
• Size class — tiny / small / medium / large
• Test coverage signal — strong / present / thin / absent
• Detected patterns — layered / hexagonal / MVC / functional core / none

Halt if size class is tiny (<500 LOC). Modernization audits are not the right tool for that scale; recommend focused code edits directly.

Phase 2: Repo state delegation

Invoke /skill:audit-existing for the descriptive baseline (what exists, what's implemented, what's missing, contracts). Carry findings forward as context for Phases 3–5. Do not duplicate its work.

Phase 3: Codebase archaeology (sample-read)

Per the rubric §3, sample-read by signal category — do not read every file:

• Breadth — first file in every top-level directory
• Depth — top 5 modules by LOC, fully
• Boundaries — every index.{ts,js,py} / __init__.py / mod.rs / equivalent
• Public surface — every file re-exported from the package entry point
• Risk hotspots — files >500 LOC; files matching utils* / helpers* / common* / lib* / misc* (kitchen-sink heuristic)
• Test parity — spot-check 3 source files for matching tests

Stop reading when the token budget hits 30% of context — leave room for synthesis.

For each read, evaluate against the rubric §4 per-theme finding patterns.

Phase 4: "What's now feasible" overlay

For each finding from Phase 3, classify the historical blocker:

• Was AI capability the blocker? (multi-file consistent rename, type-system surgery at scale, cross-module data-flow refactor) → high-confidence modernization move
• Was a project-specific constraint the blocker? (a deliberate choice to defer, a framework limitation, a contract that hasn't moved) → keep available, lower priority

This is the "AI time" overlay. It does not make taste calls — it just reweights the priority of moves whose blocker has been lifted.

Phase 5: Synthesize moves

Group findings into themes per the rubric §4: Abstraction / Separation of concerns / Types / Errors / Tests / Infra. For each move, capture:

• Theme
• Current state with file:line examples (3 max per move; representative, not exhaustive)
• Proposed state (one paragraph)
• Impact — high / medium / low
• Ease — low / medium / high (estimated effort)
• Blast radius — estimated number of files touched
• Rationale — why this is a win, in one sentence

Phase 6: Cull pass

Apply the rubric §5 cull criteria, in order:

1. Senior engineer test — would a senior engineer with no codebase context, given a clear before/after, agree the after is better? If no → cut.
2. Stylistic-preference test — objective improvement, or taste call? Cut taste calls.
3. Within-stack test — stays inside the current platform/stack? Items that fail are demoted to the out-of-scope appendix, not deleted.
4. Blast-radius vs. impact test — high blast radius (>15% files) demands high impact; >25% files demands high impact AND a verifiable benchmark (perf, type safety, etc.).
5. Cap test — keep top --max-moves (default 10) by impact × ease. Remainder → appendix.

A short list of high-confidence moves beats a long list of taste calls. Cull aggressively.

Phase 7: Emit refactor plan

Write to --out (default docs/plans/modernization-<YYYYMMDD>.md). The output is a refactor PRD, not an execution plan — it does not contain a ## Parallel Execution section. /skill:execute-prd --type=refactor consumes the refactor PRD and runs /skill:parallel-optimization itself in step 6 to produce the parallel metadata for the eventual execution plan. Format must be consumable by /skill:execute-prd --type=refactor:

---
slug: modernization-<YYYYMMDD>
source_prd: (this report)
intent: refactor
type: refactor
---

# Modernization Plan — <YYYY-MM-DD>

**Codebase:** <path>
**Shape:** <type>, <size class>, <language>
**Moves:** <N>
**Generated by:** /skill:modernize

## Summary

| Theme | Moves | Top finding |
|---|---|---|
| Abstraction | N | <one-line> |
| ... | | |

## Closed Decisions

(Inherit from /skill:audit-existing, plus modernization-specific scope ceiling: stay within current stack, no framework/API/language changes.)

## Milestone: Modernization

### Task 1: <move title>

**Theme:** Abstraction
**Current state:**
- `src/lib/foo.ts:42` — <excerpt or paraphrase>
- `src/lib/foo.ts:118` — <excerpt or paraphrase>

**Proposed state:** <one paragraph>

**Acceptance:**
- <mechanically verifiable criterion 1>
- <mechanically verifiable criterion 2>

**Estimated blast radius:** <N files>

(repeat per move; refactor task shape from /skill:execute-prd applies — characterization tests first per `--type=refactor` rules)

## Out-of-scope findings (appendix)

(Items the cull moved here. Not part of execution. Examples: framework change wanted, public API redesign signal, language switch.)

## Considered, not in primary plan (appendix)

(Items below the --max-moves cap.)

Phase 8: Handoff

If --report-only, stop after Phase 7 with:

Modernization plan written to <path>.
Run `/skill:execute-prd --type=refactor <path>` or `/batch <path>` when ready.

If --batch (or --handoff with move count > 5 or any blast-radius > 15 files), use /batch:

/batch <plan-path> [pass-through flags]

/batch spawns one isolated git worktree agent per move, runs tests per unit, and opens individual PRs. This distributes risk: a failing move doesn't block the others, and each PR is reviewable independently. Requires all moves in the plan to be independently executable (no shared mutable state between moves).

Otherwise, if --handoff (default), invoke:

/skill:execute-prd --type=refactor <plan-path> [pass-through flags]

/skill:execute-prd then runs its readiness gate, audit, validation, and execution against the modernization plan. The refactor plan-shape rules (characterization tests first) apply.

Relationship to native skills

This skill is a project-tailored modernization orchestrator unique to savviety-skills. It composes /skill:audit-existing (descriptive state) with _internal/modernization-rubric (calibration) to produce a refactor plan that feeds /skill:execute-prd --type=refactor (execution).

For ad-hoc single-file code simplification, make focused edits directly; /skill:simplify explains reports and does not refactor code. For dependency-driven modernization, use /skill:dep-audit + /skill:dep-migrate. For broader project-skills/plugin-ecosystem audit, use /skill:skill-audit. Native superpowers:writing-plans is the lighter alternative when you want to brainstorm refactor priorities without a structured rubric — but it produces a generic plan, not one calibrated to the AI-time-refresh framing.

Key Rules

1. Stay within stack. No framework, language, or public-API changes. Out-of-scope findings go in the appendix.
2. Recommend, then execute. Default to handoff; never auto-apply changes outside /skill:execute-prd.
3. Cull aggressively. A short list of high-confidence moves beats a long list of taste calls. The appendix exists so cuts aren't losses.
4. Project shape first. Hexagonal architecture is wrong for a 200-line CLI. Calibrate to detected shape before recommending.
5. Token discipline. Sample-read by the rubric's signal categories; don't read every file. Halt sampling at 30% context.
6. AI provenance is unreliable. Don't try to distinguish "old AI wrote this" from "rushed human wrote this." Evaluate code on quality terms.
7. No taste calls. Only emit moves where a senior engineer with no codebase context would agree it's a win.
8. Refactor task shape applies. Each move's acceptance must be mechanically verifiable; characterization tests come first per /skill:execute-prd --type=refactor.

Things you must not do

• Do not skip the cull pass. Long lists of low-confidence moves are the failure mode this skill exists to prevent.
• Do not invoke /skill:execute-prd until Phase 7 has written a complete plan file.
• Do not propose framework, language, or public-API changes inside primary moves. Surface them in the out-of-scope appendix.
• Do not exhaustive-read large codebases. Sample by the rubric's signal categories or halt with a "too large for safe modernization audit" finding.

Contract

• Inputs: codebase path (default cwd); optional --max-moves (default 10), --out (default docs/plans/modernization-<YYYYMMDD>.md), --handoff (default) / --report-only. Calls /skill:audit-existing, _internal/modernization-rubric, and on --handoff invokes /skill:execute-prd --type=refactor.
• Preconditions: codebase is readable; CLAUDE.md present; size is within "safe to audit" (sample by the rubric's signal categories — refuse exhaustive reads on very large repos).
• Outputs: refactor PRD at --out consumable by /skill:execute-prd --type=refactor. The PRD does not include ## Parallel Execution (that is added downstream by /skill:execute-prd step 6 via /skill:parallel-optimization).
• Postconditions: stay-within-stack constraint enforced (no framework / language / public-API changes in primary moves); out-of-scope items in the appendix; on --handoff, /skill:execute-prd takes over.
• Failure modes: codebase too large for safe sampling → halt with "too large for safe modernization audit" finding; rubric calibration impossible (no clear signal categories detected) → halt and surface the gap; never auto-apply changes outside /skill:execute-prd.