How do I install the Drain skill?

Run npx skillmds@latest add mrveiss/drain in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Drain safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Drain?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Drain free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

mrveiss (@mrveiss) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Drain by mrveiss — Agent Skill

Name: Drain
Author: mrveiss

Drain

Pick and solve the backlog issues that need no decision. Use when asked to work the backlog, drain issues, "solve what you can", run the loop autonomously, or whenever a loop tick has no assigned issue. Selects work; delegates execution to batch-implement.

mrveiss e9c397e 3.2 KB Updated Sep 17, 2026

File contents

/drain — spend the loop's attention on what it can actually finish

The loop's scarce resource is attention, not filing capacity. Picking the oldest open issue blindly burns a tick whenever it turns out to need an answer only the owner can give. This selects work that is finishable now.

Selection only. Execution belongs to batch-implement; PR mechanics to ci-pipelining.md. Do not reimplement either here.

1. Land what is already finished, before starting anything

Two queues hold finished work. gh pr list shows one; branches that never reached a PR are invisible to it and are usually the larger pile.

gh pr list --state open --json number,title --jq '.[]|"\(.number) \(.title)"'
~/.claude/scripts/drain-parked.sh                # branches whose commits never landed
~/.claude/scripts/worktree-cap.sh status         # headroom before a new worktree is refused
~/.claude/scripts/backlog-governor.sh status     # advisory signal only

Sweep → review → merge green → close with evidence → remove worktree/branch. A merged PR closes an issue; a new issue closes nothing.

Either queue non-empty → draining it IS the tick. Land the top row of drain-parked.sh before selecting new work. Zero worktree headroom makes this mandatory: the next git worktree add is refused until something lands.

2. Select

~/.claude/scripts/backlog-next.py -n 10 --why

Ranks problems → enhancements → features, FIFO within each. Judged from issue content, not labels — labels do not carry this signal (1 of 400 open issues carries needs-decision). An issue is excluded when it is an umbrella, requests a decision, has an open depends on #N blocker, or states no acceptance criteria.

Take the head of the list that does not collide with an in-flight PR's files. Skipping for collision is a deferral, not a reorder.

3. Confirm readiness before starting

The picker is a heuristic. Read the issue and abandon it if:

the acceptance criteria are not actually checkable
it needs a product or architecture call
delivering it means choosing between two reasonable designs

Put it back with a needs-decision label and options + a recommendation posted on the issue, then take the next candidate in the same tick. Never end a tick because one issue turned out to be decision-gated.

4. Execute

Hand the chosen numbers to batch-implement, which owns worktree → implement → review → merge → close → cleanup.

Hard rules

Never file issues as busywork. Filing is never blocked, but a filed issue is not progress. If nothing is solvable, say so and stop — do not manufacture backlog to look busy.
Discovered problems still get filed (that rule is absolute), just never as a substitute for solving something.
A decision-gated issue is posted and skipped, never waited on.
Evidence before done — batch-implement and the closure gate own this; do not claim a close without the verifying output.

---name: draindescription: Pick and solve the backlog issues that need no decision. Use when asked to work the backlog, drain issues, "solve what you can", run the loop autonomously, or whenever a loop tick has no assigned issue. Selects work; delegates execution to batch-implement.---# /drain — spend the loop's attention on what it can actually finishThe loop's scarce resource is **attention**, not filing capacity. Picking theoldest open issue blindly burns a tick whenever it turns out to need an answeronly the owner can give. This selects work that is finishable now.Selection only. Execution belongs to `batch-implement`; PR mechanics to`ci-pipelining.md`. Do not reimplement either here.## 1. Land what is already finished, before starting anythingTwo queues hold finished work. `gh pr list` shows one; branches that neverreached a PR are invisible to it and are usually the larger pile.```bashgh pr list --state open --json number,title --jq '.[]|"\(.number) \(.title)"'~/.claude/scripts/drain-parked.sh # branches whose commits never landed~/.claude/scripts/worktree-cap.sh status # headroom before a new worktree is refused~/.claude/scripts/backlog-governor.sh status # advisory signal only```Sweep → review → merge green → close with evidence → remove worktree/branch.A merged PR closes an issue; a new issue closes nothing.**Either queue non-empty → draining it IS the tick.** Land the top row of`drain-parked.sh` before selecting new work. Zero worktree headroom makes thismandatory: the next `git worktree add` is refused until something lands.## 2. Select```bash~/.claude/scripts/backlog-next.py -n 10 --why```Ranks problems → enhancements → features, FIFO within each. Judged from issuecontent, not labels — labels do not carry this signal (1 of 400 open issuescarries `needs-decision`). An issue is excluded when it is an umbrella, requestsa decision, has an open `depends on #N` blocker, or states no acceptancecriteria.Take the head of the list that does not collide with an in-flight PR's files.Skipping for collision is a deferral, not a reorder.## 3. Confirm readiness before startingThe picker is a heuristic. Read the issue and abandon it if:- the acceptance criteria are not actually checkable- it needs a product or architecture call- delivering it means choosing between two reasonable designsPut it back with a `needs-decision` label and options + a recommendation postedon the issue, then take the next candidate **in the same tick**. Never end atick because one issue turned out to be decision-gated.## 4. ExecuteHand the chosen numbers to `batch-implement`, which owns worktree → implement →review → merge → close → cleanup.## Hard rules- **Never file issues as busywork.** Filing is never blocked, but a filed issue is not progress. If nothing is solvable, say so and stop — do not manufacture backlog to look busy.- **Discovered problems still get filed** (that rule is absolute), just never as a substitute for solving something.- **A decision-gated issue is posted and skipped, never waited on.**- **Evidence before done** — `batch-implement` and the closure gate own this; do not claim a close without the verifying output.

Frequently asked questions