Long-Horizon Prompting
Designs auditable launch briefs for autonomous agents and parallel multi-agent work on difficult, long-running tasks.
The skill explicitly separates prompt advice from runtime-enforced budgets, permissions, logs, rollback, and approval boundaries, and includes contamination guards. It does not specify user confirmation, sensitive-data handling, rollback procedures, or transparent data-flow rules in enough detail, and several safeguards remain advisory prompt text, so points are deducted.
The material is internally structured around success predicates, failure modes, audits, stop conditions, and artifact-based reporting. However, Block 6 both permits a fallback partial report and later forbids partial returns, and there are no executable key-path tests or detailed abnormal-input failure messages. Static calibration therefore warrants a conservative score below 10.
Activation scenarios, non-fit boundaries, and adjacent-skill ownership are unusually explicit, and the template generalizes across mathematics and engineering. Adaptation for Chinese users, mainland-China network conditions, model variation, and long-run resource limits is not explicit, while some examples assume platform-specific multi-agent capabilities.
The selected skill has strong information architecture, progressive disclosure, templates, examples, gotchas, references, license metadata, and installation guidance. It lacks a clear skill version, changelog, named maintenance responsibility, and update path; the supplied repository context also contains inconsistent skill-count claims.
The reusable brief template, workflow, audit checklist, and examples can directly support the core task of designing long-horizon prompts. Static evidence provides no executed results or verified user-task outputs, and important operational controls such as budgets, permissions, and durable state are delegated to other skills or the harness.
The files provide dated research references, paper identifiers, vendor-source references, and candid caveats about the unreviewed CDC proof. Static review supplies no reproducible execution, independent verification chain, or dedicated tests covering the skill's key behavior, so evidence coverage remains limited.
- This is a low-confidence static review; the skill, examples, and CI were not executed.
- Do not treat solvability framing, effort floors, or prompt-stated permissions as real safety controls; enforce budgets, permissions, approvals, logs, and stopping conditions in an external harness.
- Block 6 contains a conflict between its partial-result fallback and its later prohibition on partial returns; clarify that fallback as external-budget exhaustion only.
- The cited research, vendor guidance, and CDC candidate proof require separate verification and should not be treated as established effectiveness or mathematical validation.
What does this skill do, and when should you use it?
This skill helps write, improve, and evaluate launch prompts for autonomous agents expected to run for hours or days. Its central method is a pseudo-formal task brief that defines terms, states an exact success predicate, excludes answer-shaped near misses, and gates return on adversarial review. It also covers portfolio diversity, blocked-route bookkeeping, evidence-based reporting, effort floors, and contamination guards. It is useful for rigorous open-ended work, but it does not provide the runtime harness, evaluator, memory system, or agent topology.
It guides the author to define load-bearing terms and edge cases, write a quantified success predicate, enumerate non-counting outcomes, list domain-specific failure modes, and specify orchestration heuristics for parallel workers. It produces reusable brief structures, pre-launch review questions, reporting contracts, return conditions, and prompt examples. It does not execute agent loops, schedule workers, enforce budgets, or replace evaluation, memory, or multi-agent infrastructure.
- A researcher is preparing a multi-hour or multi-day investigation and needs to turn “solve X” into a checkable deliverable.
- An engineering team is writing a root-orchestrator prompt and needs to preserve independent parallel approaches instead of premature convergence.
- An evaluator is reviewing a long-horizon prompt for circular reasoning, narrowed-scope answers, fabricated progress, and other plausible near misses.
- A team is diagnosing a failed autonomous run that returned a status report or answer-shaped artifact instead of a verified result.
- A technical lead needs effort floors, evidence-traceable claims, external-search boundaries, and an audit-gated return condition in a complex task brief.
What are this skill's strengths and limitations?
- Provides a complete brief anatomy from definitions and success predicates through adversarial return gates.
- Explicitly addresses narrowed-scope answers, unproved reductions, bounded verification, circular satisfaction, and status-report theater.
- Offers concrete diversity policies including approach-family registries, blocked-route bookkeeping, early independence, and late cross-pollination.
- Separates prompt-level guidance from constraints that should be enforced by the runtime harness.
- It designs launch prompts rather than executing loops, coordinating workers, maintaining ledgers, or enforcing permissions and budgets.
- It depends on domain experts to supply the relevant edge cases and failure modes.
- The cited Cycle Double Cover candidate proof had no independent peer review or formalization at publication, so the example does not establish the theorem’s correctness.
- It does not provide platform-specific API configuration or cost estimates.
How do you install this skill?
Copy the complete skills/long-horizon-prompting directory from the repository into an Agent Skills-compatible directory, such as .codex/skills/ for a Codex project or .agents/skills/ for a generic project. Do not flatten the skill into a single SKILL.md file because the repository uses a directory layout. The README also documents Claude Code marketplace installation: /plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering, followed by installing context-engineering@context-engineering-marketplace.
How do you use this skill?
Activate it when writing or reviewing a long-running agent prompt. For example: “Write a launch brief for a long-running parallel research task with an exact success predicate, non-counting outcomes, domain failure modes, portfolio-diversity rules, evidence-based reporting, and an adversarial return gate.” The skill then structures the work around the predicate, definitions, audit checklist, orchestration policy, and return condition. Specific runtime commands, budget enforcement, and external-tool configuration are not defined by this skill.
How does this skill compare with similar options?
Compared with multi-agent-patterns, this skill owns the wording and orchestration policy of the prompt, not topology, handoffs, or coordination protocols. Compared with harness-engineering, evaluation, and advanced-evaluation, it specifies prompt-level constraints and audit requirements rather than runtime enforcement, deterministic evaluators, or judge design.