Skip to content
Back to skills

Ashlr Cost Refactor

ASecurity

Token-efficient refactoring discipline for Claude Code. Diff-first before editing, structural edits for renames, find-all-refs then one multi_edit per file, no re-reads between edits, one genome proposal at the end. Reduces refactor session cost 40–65%.

  • 3 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 29, 2026
ai-agentsbashrefactoring

Works with

  • claude code

Security analysis

A100/100

Scanned September 29, 2026

npx -y skills add ashlrai/ashlr-plugin --skill ashlr-cost-refactor --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ashlr Cost Refactor?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ashlr Cost Refactor
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/ashlrai-ashlr-cost-refactor/badge)](https://www.skillsdirectory.com/skills/ashlrai-ashlr-cost-refactor)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ashlr-cost-refactor
description: Token-efficient refactoring discipline for Claude Code. Diff-first before editing, structural edits for renames, find-all-refs then one multi_edit per file, no re-reads between edits, one genome proposal at the end. Reduces refactor session cost 40–65%.
---

# Ashlr Cost Refactor

A structured discipline for refactoring tasks that minimizes token spend without sacrificing correctness.

## Rules (active when this skill is enabled)

### 1. Diff first
Before editing anything, run `ashlr__diff` or `ashlr__diff_semantic` to understand what already changed (staged/unstaged). This prevents redundant edits to files that are already in the state you want.

### 2. Structural edits for renames
When renaming a symbol, type, or file — use `ashlr__edit_structural` instead of search/replace. Structural edit finds all references in one pass and renames them correctly, including imports and re-exports.

### 3. Find all refs, then one `ashlr__multi_edit` per file
Before editing a file, run `ashlr__grep` to find all references. Then apply all changes to that file in a single `ashlr__multi_edit` call. Do not make sequential single edits to the same file — each one burns a full round-trip.

### 4. No re-reads between edits in a sequence
Once you've read a file and know its structure, proceed through your edit sequence without re-reading. `ashlr__edit` and `ashlr__multi_edit` return diffs that confirm each change — that's sufficient to continue.

### 5. One `ashlr__genome_propose` at the end
After the refactor is complete and stable, write one genome proposal summarizing the entire change: what was renamed/moved, why, and the affected file paths. Do not propose mid-refactor.

### 6. Verify with `ashlr__bash` not `Read`
After a refactor, confirm with `bun run typecheck` or `bun test` via `ashlr__bash`. Do not re-read edited files to verify — the typecheck output is authoritative and far cheaper.

## Sequence template

```
1. ashlr__diff_semantic()             — see what's already changed
2. ashlr__grep("OldName|oldFn")       — find all refs
3. ashlr__edit_structural(...)        — rename symbols
   OR ashlr__multi_edit([...])        — batch all edits per file
4. ashlr__bash("bun run typecheck")   — verify, not Read
5. ashlr__genome_propose(summary)     — one proposal at the end
```

## Anti-patterns

- Running `Read` before each edit in a multi-file refactor
- Making 3 sequential `ashlr__edit` calls on the same file
- Re-reading a file after editing it to confirm the change
- Calling `ashlr__genome_propose` after each individual edit

## Activation

Enable via `/ashlr-cost-refactor on` or by writing `{"enabled":true}` to `~/.ashlr/cost-refactor.json`. Toggle off with `/ashlr-cost-refactor off`.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.