Sift — Git skill for Claude Code
A context budget for CI.
How to install Sift
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open Jithendrag22/sift and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Sift does
A context budget for CI. The bundle-size bot, for agent configuration — measures the always-on context tax of your CLAUDE.md, skills and subagents, diffs it per pull request, and fails the build when it grows.
Alternatives in Git
- Worktree Deliver — Commit all work, push, and create a pull request from the current worktree 23.4k ★
- Yeet — Stage, commit, push code, and open a GitHub pull request via CLI 14.6k ★
- Generate PR Description — You are tasked with generating a comprehensive pull request description following the repository's standard te 10k ★
README
Sift
**A context budget for CI.** The bundle-size bot, for agent configuration.
[](https://github.com/Jithendrag22/sift/actions/workflows/tests.yml) [](LICENSE) [](#requirements) [](#requirements)
Your `CLAUDE.md`, your skills, your subagents load into **every session, for every engineer, on every task**. Nobody is watching the total.
Sift puts the number on the pull request that moved it.
### ❌ Context budget — FAIL
████████████████████████████ 2,534 / 1,000 always-on tokens (253%)
+2,469 always-on tokens vs base.
That prefix is cache-written once at 2× and re-read at 0.1× on each of ~45 turns,
so it bills as +15,801 input tokens per session (×6.4).
| file | kind | was | now | Δ |
|--------------------------------------------|--------------|-----:|-----:|-------:|
| CLAUDE.md | instructions | 23 | 1908 | +1,885 |
| .claude/skills/changelog-writer/SKILL.md | skill | 0 | 118 | +118 |
Why
Agent configuration grows the way dependency trees grow: one reasonable addition at a time, each individually defensible, nobody watching the total.
JavaScript did not solve bundle bloat by being clever. It solved it **socially** — by putting the number on the pull request that caused it, while someone could still change their mind. Agent configuration has no such number.
And the cost is not only money:
| source | finding |
|---|---|
| Anthropic, Tool Search Tool | Cutting always-loaded tool definitions from ~77K to ~8.7K tokens raised MCP accuracy 49% → 74% (Opus 4) and **79.5 |
Related Skills
Daedalai Sync Claude Md
Reconcile the project's local CLAUDE.md with its canonical CLAUDE_MD document in DaedalAI — pull when remote i
Address Review Preexisting Unrelated Precommit Failure
Use when a local pre-merge/pre-push gate goes red on work your PR did not introduce. Covers two verified varia
Read PR
Executes 'pr-context-retriever' skill and accomplishes the user request using that skill.
Peqy
Review peqy bot comments on a PR and assess their validity against the codebase
Consolidate Issues
Review every open issue against git history, close what's already done or stale, migrate the legacy ones, slim
Build Bundle
Build one governed template bundle end to end, from research fan-out through the four-lens review to a merged
Related Agents
PR Open
Create a pull request for the current branch. Analyzes all commits, runs validation, and opens a well-formatte
PR Security Review
Use this agent to perform a security analysis of a pull request BEFORE checking out or running any code locall
PR Change Analyzer
Analyzes the changes a pull request makes to a single module. Receives module source, diffs, and PR context. R