Guide · Agent skill
How to use the Long-Horizon Prompt Writer agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The long-horizon-prompting skill is an agent skill that helps you write or review the launch prompt for an agent that will work autonomously for hours or days, alone or as an orchestrator of many parallel sub-agents. It turns a vague hard problem into a pseudo-formal task brief with defined terms, an exact success test, a list of outcomes that do not count, and rules for when to continue, stop or return. This guide explains what the skill does, how to install it, how to trigger it and where its limits are.
Long-Horizon Prompt Writermuratcankoylan/Agent-Skills-for-Context-Engineering · MIT Download the skill Quick start
- 1Download the ZIP. Get the long-horizon-prompting skill ZIP from the directory page. The package includes SKILL.md and a references/ folder.
- 2Unzip into a skills folder. Extract it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder for one project.
- 3Or upload in the Claude apps. In the Claude apps, open Settings, go to Capabilities, then Skills, and upload the ZIP.
- 4Describe the long run. Ask for help with the launch prompt, stating the problem, the expected duration and whether parallel sub-agents are involved.
- 5Review the brief before launch. Check the definitions, success test, non-counting outcomes and stop conditions the agent proposes, and add any missing machine-checkable tests.
What is an agent skill and what does this one do?
An agent skill is a folder of instructions that an AI agent such as Claude can load when a task matches the skill's description. This skill is written for one situation: preparing the prompt that launches a long-running autonomous task, or an orchestrator that manages many parallel workers on an open-ended problem.
Once it is active, the agent structures that launch prompt as a precise task brief. The brief defines key terms (including edge cases a lazy solution could exploit), states one exact success predicate, lists non-counting outcomes such as partial progress and best-effort summaries, sets persistence rules with stop and return conditions and effort floors, and requires an adversarial audit before the result is accepted. It also points you to other skills for agent architecture, evaluation, judge design and memory.
- Definitions with degenerate cases, defined before the goal
- One exact success predicate for the returned artifact
- An enumerated list of non-counting outcomes
- Persistence rules: explicit stop and return conditions and effort floors
- An approach registry and blocked-route bookkeeping for parallel work
- A return gated on adversarial audit
Install the long-horizon-prompting skill
Download the skill ZIP from the directory page. Unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only in that project. The skill's folder should contain SKILL.md and a references/ folder.
If you use the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/ (available across your projects)
- Project install: .claude/skills/ inside the project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
How do you trigger the skill with a natural request?
You do not need a special command. The skill activates when your request matches its description: writing, enhancing or evaluating the launch prompt for a long autonomous run or a multi-agent orchestration on a hard problem. Say what you are about to launch and what kind of problem it is.
Requests that tend to match the skill's purpose include asking the agent to review a prompt before you start a multi-day run, to add stop conditions to an orchestrator prompt, or to explain why a previous run returned early. If you want the skill applied, name the task and its length.
- Write a launch prompt for an autonomous run that will take several hours on this problem: [problem]
- Review this orchestrator prompt that manages 20 parallel workers. Add stop conditions and an audit gate.
- My last long agent run returned a near-miss answer and claimed it was finished. Diagnose the brief.
- Turn this vague goal into a task brief with a success predicate and non-counting outcomes.
What you get back and what it does not do
The skill shapes the instructions. It does not run the long task, provide an execution environment or act as an evaluator. You still need an agent environment to run the work and a way to score the result.
Machine-checkable parts should be checked by commands rather than prose. The skill's own advice is to express each such part as a named check with the command that runs it, and to keep the linguistic predicate only for what no command can decide. Simple, short or interactive prompts are poor candidates, because the extra structure adds cost without benefit.
- Does not execute the long task or provide a sandbox
- Does not build evaluators or quality gates; route those to the evaluation skill
- Routes architecture, runtime control and memory questions to related skills in its description
Safety notes: scripts and network access
The skill lists no requirements. According to its listing, it does not run scripts and does not use the network. Its files are the SKILL.md instructions and reference documents in references/, which cover the task-brief template, vendor guidance, handoff contracts, research evidence and an annotated example prompt.
Because it only shapes text, review the brief yourself before launching a costly run. Long runs spend real compute, and the skill itself notes that a brief with a loophole can produce an answer-shaped artifact that does not solve the problem.
- Runs scripts: no
- Uses network: no
- Requires: nothing listed
- Licence: MIT
Tips for better briefs
Start with the success predicate and the non-counting outcomes, since the skill treats the excluded outcomes as the highest-leverage part of the brief. Then add an enumerated list of the specific ways a result could be subtly wrong, so the auditor has a concrete hunt list.
For orchestrated runs, specify what each worker must return and how the orchestrator checks it, and keep a registry of approaches already tried so parallel workers do not converge on one route.
- Define terms, including degenerate cases, before stating the goal
- List outcomes that do not count, such as partial progress and bounded checks
- Name the failure modes the auditor should hunt for
- Use the approach registry and blocked-route notes for parallel workers
- Pick a different skill when the question is about architecture, evaluation or memory
What people use it for
- Preparing a multi-hour autonomous run on a hard, open-ended research or analysis problem
- Writing an orchestrator prompt that assigns and collects results from many parallel worker agents
- Adding explicit stop conditions, effort floors and an audit gate to an existing agent prompt
- Diagnosing a previous long run that returned early, gave a near-miss answer or falsely claimed completion
Frequently asked questions
Is the long-horizon-prompting skill free?
The skill is published under the MIT licence, as stated on its listing. The source does not list a price, so check the directory page for any terms beyond the licence.
What do I need to use it?
The skill's listing says it requires nothing beyond an agent that can load skills, such as Claude with skills enabled. It does not run scripts and does not use the network.
Where do I install the skill folder?
Unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, you can upload the ZIP under Settings, Capabilities, Skills.
Will the skill run my long task for me?
No. It shapes the launch prompt and audit rules. It does not provide an execution environment or an evaluator, so you still need a way to run the work and score the result.
When should I not use this skill?
Skip it for short, simple or interactive prompts, where the extra structure adds cost without benefit. Also skip it when a test suite alone can decide success, since the skill advises machine-checking those parts with commands.
How is this different from a general prompt-writing skill?
It is narrower. It focuses on long autonomous runs and parallel orchestration, with exact success predicates, non-counting outcomes, persistence rules and adversarial audit gates, and it sends architecture, evaluation, judge design and memory questions to other skills.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are preparing a long autonomous run or an orchestrator prompt on a hard, open-ended problem and want the success test, stop rules and audit gate spelled out before you spend compute. Pick something else for short interactive prompts, for tasks a test suite can fully judge, or if you need an agent that actually executes the work.
Download the skill