How do I install the Long Task skill?

Run npx skillmds@latest add evan-daruwalla/long-task 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 Long Task 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 Long Task?

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 Long Task free to use?

Yes. Installing skills from SkillMD is free. This skill is licensed under MIT.

Who published Long Task?

Evan-Daruwalla (@evan-daruwalla) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Long Task by evan-daruwalla — Agent Skill

Name: Long Task
Author: Evan-Daruwalla

Long Task

Execution loop for work longer than one sitting or more than ~7 steps: a numbered plan where every step ends in a check, riskiest step first, a one-line state tracker at the top of each turn, an iteration budget, and a progress lock — the same command failing the same way twice means change approach or stop and report NO-PROGRESS (hooks/no-progress.js fires this deterministically on PostToolUseFailure). Invoke: "long-task", "plan this out", "this is a big one", multi-hour or multi-session work. Small work → small-task. Cheaper-tier delegation → opus-workers. Launch risks → pre-mortem.

Evan-Daruwalla c3e6757 2 files · 9.9 KB Updated Sep 17, 2026

File contents

long-task

1. Plan before the first edit

Numbered steps, each one a verifiable unit:

N. <step> → verify: <command or observation that proves it>

Riskiest or most uncertain step first — it is the one most likely to change the plan, so learn that before spending on the rest. Set a BUDGET: max tool turns per step (default 25; the user's cap if they gave one). Say the plan before executing it; a plan nobody can read is not a plan.

2. State tracker — first line of every turn while the task is open

[LONG-TASK] step k/N · done: <verified steps> · blocked: <error, or none> · next: <the single step this turn>

One line. It exists for the reader at high context, when the plan written 80k tokens ago is no longer in view — the same reason small-task and compact-io are injected per prompt.

3. The loop

Analyze the last result → pick the single next step → execute → verify with real output. Advance only on a passed check. A step that will not pass inside its budget is a finding, not a retry: report it and re-plan.

4. Progress lock

The same command producing the same failure twice in a row means you are spinning. hooks/no-progress.js injects [NO-PROGRESS] when it sees exactly that. On it: change the approach — a different command, a different diagnosis, read the code instead of re-running it — or stop and report a NO-PROGRESS BLOCK naming the command and the error. Never a third identical attempt.

5. Done

Before declaring the goal done: every step's verify ran and its real output is in the transcript; the project's own definition of done applies on top. An unmet check means the loop continues with a corrective step, or the shortfall is reported as unmet — never rounded up to done.

Hand-offs

Delegating steps to a cheaper tier → opus-workers. Launch-blocking risks before shipping → pre-mortem. Recording the run → the project's record.

Config

Off for a session: NO_PROGRESS_OFF=1.
The repeat threshold and error normalisation live at the top of hooks/no-progress.js.

---name: long-taskdescription: Execution loop for work longer than one sitting or more than ~7 steps: a numbered plan where every step ends in a check, riskiest step first, a one-line state tracker at the top of each turn, an iteration budget, and a progress lock — the same command failing the same way twice means change approach or stop and report NO-PROGRESS (hooks/no-progress.js fires this deterministically on PostToolUseFailure). Invoke: "long-task", "plan this out", "this is a big one", multi-hour or multi-session work. Small work → small-task. Cheaper-tier delegation → opus-workers. Launch risks → pre-mortem.license: MIT---# long-task## 1. Plan before the first editNumbered steps, each one a verifiable unit:```N. <step> → verify: <command or observation that proves it>```Riskiest or most uncertain step first — it is the one most likely to changethe plan, so learn that before spending on the rest. Set a BUDGET: max toolturns per step (default 25; the user's cap if they gave one). Say the planbefore executing it; a plan nobody can read is not a plan.## 2. State tracker — first line of every turn while the task is open```[LONG-TASK] step k/N · done: <verified steps> · blocked: <error, or none> · next: <the single step this turn>```One line. It exists for the reader at high context, when the plan written80k tokens ago is no longer in view — the same reason small-task andcompact-io are injected per prompt.## 3. The loopAnalyze the last result → pick the single next step → execute → verify withreal output. Advance only on a passed check. A step that will not pass insideits budget is a finding, not a retry: report it and re-plan.## 4. Progress lockThe same command producing the same failure twice in a row means you arespinning. `hooks/no-progress.js` injects `[NO-PROGRESS]` when it sees exactlythat. On it: change the approach — a different command, a differentdiagnosis, read the code instead of re-running it — or stop and report a**NO-PROGRESS BLOCK** naming the command and the error. Never a thirdidentical attempt.## 5. DoneBefore declaring the goal done: every step's verify ran and its real outputis in the transcript; the project's own definition of done applies on top.An unmet check means the loop continues with a corrective step, or theshortfall is reported as unmet — never rounded up to done.## Hand-offsDelegating steps to a cheaper tier → opus-workers. Launch-blocking risksbefore shipping → pre-mortem. Recording the run → the project's record.## Config- Off for a session: `NO_PROGRESS_OFF=1`.- The repeat threshold and error normalisation live at the top of `hooks/no-progress.js`.

Frequently asked questions