Premature return
Long trajectories drift toward uncertainty, status reporting, and early abandonment before the requested artifact exists.
Long-Horizon Prompting / v1.0
Prompts for work that does not end in one turn.
Long-horizon prompting treats the launch prompt as a specification surface: define exact success, close the likely escape hatches, require evidence, and return only when the artifact survives an adversarial check.
What it is
A pseudo-formal task brief, written in natural language.
Formal verification starts with a machine-checkable specification. Many hard research and engineering tasks do not have one. The transferable discipline is to state the acceptance condition precisely enough that an adversarial reader cannot satisfy its letter without satisfying its intent.
This is not a prompt for making a model “try harder.” It is a method for deciding what the run is allowed to call complete, what evidence must exist, and which constraints must move out of prompt text and into the runtime harness.
Why long runs fail
Additional compute amplifies the specification.
Answer-shaped near misses
A narrowed scope, lucky benchmark run, unproved reduction, or scanner dump can resemble the requested result without being it.
Verification bottlenecks
Generating more candidates helps only when the system can identify which candidate is actually correct.
Diversity collapse
Parallel workers share priors and converge quickly unless early independence and blocked-route bookkeeping are designed in.
Brief anatomy
Ten blocks with distinct jobs.
- DefinitionsFix load-bearing terms and boundary cases.
- Success predicateState exactly what must be true at return.
- Non-counting outcomesName the plausible near misses.
- Solvability framingUse only where a complete solution plausibly exists.
- Orchestration policyPreserve independent approaches in parallel search.
- Verification policyGive fresh reviewers a domain-specific attack list.
- Reporting contractRequire artifacts and evidence, not status.
- Return conditionGate completion on the artifact, not confidence.
- Effort framingRemove permission to quit after the first failed wave.
- Contamination guardsDefine what external retrieval may contribute.
System boundary
The prompt is one layer, not the whole system.
PromptSemantic success, named near misses, evidence contract, return rule.
HarnessBudgets, permissions, sandboxing, locked evaluators, rollback.
EvaluatorIndependent acceptance checks, held-out data, fresh-context review.
Durable stateProgress ledger, approach registry, blocked routes, rejected candidates.
GovernanceHuman control of evaluator changes, deployment, and promotion.