eliptikwu
Claude Code skills by EliptikWu (LK Industries).
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/eliptikwu--next-task)HTML
<a href="/listings/eliptikwu--next-task"><img src="/api/listings/eliptikwu--next-task/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
**Claude Code, as a plugin:**
/plugin marketplace add EliptikWu/next-task /plugin install next-task@eliptikwu
Or copy `skills/next-task/` to `~/.claude/skills/next-task/` (every project) or to `<repo>/.claude/skills/next-task/` (one project).
**Other tools:** copy `skills/next-task/` to `<repo>/.agents/skills/next-task/` for one project, or to `~/.agents/skills/next-task/` for all of them. Every tool below reads that folder.
| Tool | Run it | Fresh session | |---|---|---| | Codex | `$next-task` | `/new` | | Gemini CLI | say "next task" (it picks the skill itself) | `/clear` | | GitHub Copilot in VS Code | `/next-task` | New Chat (Ctrl+N) | | GitHub Copilot CLI | `/next-task` | `/clear` | | Cursor | `/next-task` | New chat (Cmd/Ctrl+N) | | Devin Desktop (formerly Windsurf) | `/next-task` | a new conversation |
The project's instructions go in whatever file your tool reads (`CLAUDE.md`, `AGENTS.md`, `GEMINI.md`...). The skill follows that file.
If a project already has its own skill for this loop, that one wins.
No backlog yet? The skill offers to create one. The format is in [`skills/next-task/reference/backlog-format.md`](skills/next-task/reference/backlog-format.md).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 16 lines
- {
- "name": "eliptikwu",
- "description": "Claude Code skills by EliptikWu (LK Industries).",
- "owner": {
- "name": "EliptikWu"
- },
- "plugins": [
- {
- "name": "next-task",
- "source": "./",
- "description": "One task per session, then /clear: measured 60 % fewer tokens per PR without skipping tests or review.",
- "version": "1.0.0",
- "license": "Apache-2.0"
- }
- ]
- }
README.md · README · 187 lines
- <img src=".claude-plugin/icon.png" width="96" alt="">
- # next-task
- [Leer en español](README.es.md)
- An agent skill for working through a backlog **one task per session**. It uses the open
- [Agent Skills](https://agentskills.io) format, so it runs in Claude Code, Codex, Gemini
- CLI, GitHub Copilot, Cursor and Devin Desktop (formerly Windsurf). Each session it
- picks the next item and takes it to a verified pull request. It leaves a one-line handoff
- in the repo, then tells you to start a fresh session (`/clear`) and type `/next-task` again.
- You type two commands. Each task costs 60 % fewer tokens, without skipping tests or review.
- ## Measured savings
- These numbers come from 561 real pull requests in two production projects:
- - [LK Creator](https://lkcreator.com), a FastAPI + Next.js SaaS that converts custom
- Minecraft content.
- - ISO 21101 turismo, a safety-management web app for adventure-tourism operators.
- The data is the token usage that Claude Code already stores in its local session
- transcripts. Same developer, same model (Claude Opus), November 2026. Other tools were not
- measured: the idea should carry over, since they all re-read the conversation on every
- step, but we have no numbers for them.
- | | One task per session | Chained tasks (≥3 PRs in one session) | Savings (95 % CI) |
- |---|---|---|---|
- | Cost per PR, LK Creator | $4.39 (162 PRs) | $14.58 (131 PRs) | **70 %** (37–72 %) |
- | Cost per PR, ISO 21101 turismo | $5.10 (49 PRs) | $9.63 (219 PRs) | **47 %** (18–78 %) |
- | **Cost per PR, both** | | | **60 %** (42–74 %) |
- | **Context per model call, both** | 123–130k tokens | 342–357k tokens | **73 %** (62–75 %) |
- **On a subscription.** We read `/usage` three times, around two work periods, and
- summed the tokens of every session in between, across all projects and including
- subagents.
- - One point of the weekly limit cost $17.4 of API-equivalent tokens: 11 points in total,
- 0.059 and 0.057 points per dollar in the two periods.
- - `/usage` rounds to whole points, so this figure is ±9 %.
- - So one task per session uses about **0.25–0.29 % of the weekly limit per PR**. Chaining
- tasks uses **0.55–0.84 %**. That is roughly **twice as many PRs per week** on the same
- plan.
- **Why it works.** Every model call re-reads the whole conversation. In a chained session
- the context per call keeps growing; it passed 500k tokens after about 10 tasks. So each
- later task pays for all the earlier ones. A fresh session starts every task with the same
- small context.
- ### Quality: no worse, and the safety steps actually happen
- We joined the first 510 of these PRs with their GitHub history:
- | | One task per session | Chained | Difference (95 % CI) |
- |---|---|---|---|
- | Merged | 98.2 % | 98.6 % | none (−3.5 to +3.1 pts) |
- | Reverted later | 0 % | 0 % | none |
- | Tests / typecheck run before opening the PR | **97.6 %** | 75.1 % | 22.5 pts more (2 to 38) |
- | Code review run before opening the PR | **87.9 %** | 14.8 % | 73 pts more (55 to 87) |
- In long chained sessions the model quietly starts skipping verification and review as the
- context fills up. A fresh session follows the full checklist almost every time.
- We also measured "rework within 7 days": a later `fix` PR that touches the same code files.
- It favored one task per session on ISO 21101 turismo (8.6 % vs 59.2 %), but it is too
- noisy to claim. In these repos a third of all PRs are planned `fix` work on the same
- modules, so the combined interval crosses zero.
- ### Behavior tests
- Before release we ran the skill and its previous version on three scripted repos, graded
- automatically (scripts in [`tests/`](tests/)):
- - **a real bug first:** fix it, one task only;
- - **an already-fixed item first:** check the premise, mark it obsolete, move on;
- - **no backlog:** offer to create one and don't invent work.
- This version passed 19/19 checks with the same token use as the previous one (≈54k per
- run). The previous version wrote invented PR numbers into the handoff when there was no
- remote, and this version fixes that.
- ### How it was measured
- The scripts are in [`bench/`](bench/).
- - **Cost.** We took the tokens of every model call, including subagents, and weighted them
- by Anthropic's price ratios: cache read 0.1×, cache write 1.25× or 2×, output 5×. We
- checked these weights against the cost Claude Code records, and they agree within ±10 %.
- - **Task.** A task is a pull request actually created in that session: a `gh pr create`
- that returned a PR URL.
- - **Confidence intervals.** We ran a 5,000-sample bootstrap over **sessions**, not tasks.
- That way one long session doesn't count as 100 independent samples.
- ### Limits, stated plainly
- - It's observational, not an A/B test. The tasks in the two groups are different, and the
- chained group has only 11 sessions.
- - The dollar figures are API-price equivalents. The subscription figure comes from two
- periods and should be read as an estimate.
- - The quality table shows that checks were *run*. It cannot show that each one was run
- well.
- ### Run it on your own Claude Code sessions
- ```bash
- python bench/measure.py ~/.claude/projects/<your-project-folder> next-task sessions.json
- python bench/compare.py sessions.json
- python bench/bootstrap.py "mine=sessions.json"
- ```
- ## What it does in each session
- 0. **Guard.**
- - If this conversation already closed a task, it asks you to `/clear` first.
- - If one of your PRs is still open, it finishes that one first.
- - If the default branch's CI is red, fixing it *is* the task.
- 1. **Pick.** It takes the first open item of the backlog (`BACKLOG.md`, `TODO.md`, or
- GitHub issues), reading only the part it needs.
- 2. **Check the premise.** It confirms in the code that the problem still exists, and drops
- items that are already fixed.
- 3. **Do it.** It finds the root cause and fixes it in the shared mechanism, in the
- surrounding style.
- 4. **Verify.** It runs the checks CI would run, then reviews the diff.
- 5. **Open the PR.** It merges only if the project's instructions file explicitly allows it, and only when
- checks pass.
- 6. **Hand off.** It writes one line in the backlog, with the PR number and the one thing
- the next session must know.
- 7. **Stop.** It tells you to start a fresh session (`/clear`), then `/next-task`.
- It follows these token rules throughout:
- - the repo is the memory;
- - it reads line ranges, not whole files;
- - wide searches go to a subagent that answers in under 15 lines;
- - every command's output is trimmed;
- - it uses CLI tools over MCP;
- - it never re-derives a fact it already established;
- - it stops after two failed attempts at the same fix;
- - compaction is treated as a sign that the task was too big.
- It works without a GitHub remote. In that case it commits on a branch and leaves the PR to
- you.
- It answers in your language; the skill itself is written in English because that's what
- the model reads.
- ## Install
- **Claude Code, as a plugin:**
- ```
- /plugin marketplace add EliptikWu/next-task
- /plugin install next-task@eliptikwu
- ```
- Or copy `skills/next-task/` to `~/.claude/skills/next-task/` (every project) or to
- `<repo>/.claude/skills/next-task/` (one project).
- **Other tools:** copy `skills/next-task/` to `<repo>/.agents/skills/next-task/` for one
- project, or to `~/.agents/skills/next-task/` for all of them. Every tool below reads that
- folder.
- | Tool | Run it | Fresh session |
- |---|---|---|
- | Codex | `$next-task` | `/new` |
- | Gemini CLI | say "next task" (it picks the skill itself) | `/clear` |
- | GitHub Copilot in VS Code | `/next-task` | New Chat (Ctrl+N) |
- | GitHub Copilot CLI | `/next-task` | `/clear` |
- | Cursor | `/next-task` | New chat (Cmd/Ctrl+N) |
- | Devin Desktop (formerly Windsurf) | `/next-task` | a new conversation |
- The project's instructions go in whatever file your tool reads (`CLAUDE.md`, `AGENTS.md`,
- `GEMINI.md`...). The skill follows that file.
- If a project already has its own skill for this loop, that one wins.
- No backlog yet? The skill offers to create one. The format is in
- [`skills/next-task/reference/backlog-format.md`](skills/next-task/reference/backlog-format.md).
- ## License
- [Apache 2.0](LICENSE). You can use, change and share it. Copies and derived works must
- keep the copyright and the [NOTICE](NOTICE) file, and must say what was changed. The
- license doesn't grant use of the author's name to promote a derived work.
- Made by [EliptikWu](https://github.com/EliptikWu) at LK Industries.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.