Intent-framed delegation is a control regime; prompt length does not establish it

Type: kb/types/note.md

In a consequential planner--executor handoff, intent-framed delegation is a governed allocation of information and decision rights. Upstream fixes purpose and binding boundaries that standing doctrine cannot determine for the task. A competent, authorized executor may select or construct means when execution can expose evidence that could change which permitted means serve that purpose. Prompt text is only one carrier in this arrangement. Its length does not establish the control regime, although a short handoff can participate in one when a verified surrounding system supplies the missing controls.

Reciprocal opposition explains the military source problem

The military problem is deeper than ordinary distributed information. The U.S. Marine Corps' MCDP 1 formulation describes an opponent as an independent actor with its own objectives and plans that resists and acts in return. It also directly describes war as uncertain, dynamic, disorderly, and dominated by friction. Chance and information unavailable in advance can also defeat advance specification without an adversary; that boundary is synthesis here, not a direct MCDP attribution. These conditions overlap but are not the same. Generic friction can disrupt a non-adversarial process. Reciprocal opposition makes the uncertainty interactive because another purpose-bearing actor helps produce the future state.

The next step is a cross-source synthesis, not wording supplied by one source. When an opponent can observe and respond, one side's action can change the opponent's next choice, which changes the state for later action. Some action-to-outcome relations therefore depend on another chooser. Friction and information limits add further divergence. The record does not show that opposition was the sole cause of Auftragstaktik or establish its weight relative to communications, scale, technology, training, or culture.

These conditions make the source response intelligible. Complete advance control is unavailable. Some facts that discriminate among possible means become visible only near the point of action. Purpose can remain useful after a task or method becomes obsolete. MCDP 1 therefore assigns a mission and purpose without prescribing every means, while retaining guidance, authority limits, coordination, responsibility, and reporting. The U.S. Army's ADRP 6-0 formulation likewise joins intent and desired results with point-of-action initiative, resources, specified authority, coordination, supervision, accountability, and retained upstream responsibility. The generalized chain from interactive uncertainty to later-informed, bounded choice is cross-source synthesis; the individual doctrinal elements are direct source claims.

Stahel's historical account describes Prussian-German thought as treating war as contingent and leadership as direction rather than complete control. It also rejects one generally valid, uniform historical Auftragstaktik and treats it as distinct from modern mission command. MCDP 1 and ADRP 6-0 can state coherent modern prescriptions on their own terms, but they do not establish identity with historical German practice, continuity, or effectiveness.

Transfer requires the later-informed choice mechanism

A causal-origin match means the target shares the condition that generated the source problem. A downstream mechanism match means the target lacks that origin but independently has the later-informed choice problem to which bounded discretion responds. Because borrowed patterns transfer only over a shared mechanism, source warrant reaches only the match the target actually has. Neither kind of match establishes target effectiveness.

Genuinely adversarial agent work may share the causal origin when another capable actor has independent goals, observes or responds to the agent's actions, and deliberately changes the state or available options. That is a Commonplace inference from the military premises, not an agent result. Non-adversarial work does not need hostile purpose. It shares the downstream mechanism only when it has an independent route to later-arriving, means-relevant information:

  • a bounded choice among permitted means can affect execution;
  • upstream exposes purpose, binding constraints, acceptance conditions, privileged facts, external commitments, and relevant coupling that execution cannot safely recover;
  • a named observation or produced result can arrive during execution, and at least one possible result would change selection, timing, modification, or abandonment of a permitted means;
  • the executor can access that evidence and is competent and authorized for the choice; and
  • the choice remains within authority and coordination bounds, with common selection or coordination-bearing ownership where local choices could conflict.

Here, later-arriving means unavailable or unsettled at the relevant upstream specification point, not merely read later. Means-relevant means capable of changing the bounded choice. The more general allocation rule is to fix what the executor cannot determine, not what it will. These are conditions for this transfer argument, not an experimentally proven minimum or a sufficient implementation packet.

Information and decision rights run in both directions

Upstream must preserve what live execution cannot recover: intent, cross-task coupling, privileged facts, constraints, external commitments, output ownership, and done conditions. The executor uses current state, tool results, local failures, newly exposed constraints, and evidence produced during execution when selecting authorized means. The executor need not know more overall; its information advantage is specific to the delegated decision.

The open choice remains answerable to the fixed purpose and boundaries. The executor must be able to derive feasible means, remain within authority, and detect when evidence or authority is insufficient. Delegation changes who holds that judgment. Deferral changes when it is made. Delegating immediately to an actor that already has the evidence is delegation without a later information advantage.

Shared doctrine and task intent complete different omissions

In the source domain, MCDP 1 says that its command philosophy applies during preparation and requires competent leaders, familiarity, and shared understanding. ADRP 6-0 joins shared understanding and clear intent to local adaptation, and says that applying its principles requires education, rigorous training, and experience. The sources also say that intent can continue to guide action when a task or prior order no longer fits. These are claims about the source systems. The omission classes below are a Commonplace synthesis, not military terminology or an attribution of the agent method to those sources.

When a verified consumption path gives an executor Commonplace doctrine, that standing instruction supplies a shared way to interpret work: purposes, vocabulary, selected methods, defaults, precedence, and authority rules. It can recover an answer already selected by the system, but it can also guide generation of a means that no upstream actor selected. In the second case, task intent supplies the task-specific purpose and success criterion, binding constraints and commitments delimit acceptable adaptation, and execution evidence selects among or helps construct permitted means. Reconstructing an inherited answer and deriving a new means are therefore different kinds of completion.

For each consequential unstated choice, the executor must be able to identify one of four cases:

  • Inherited: verified doctrine or a contract governs the choice.
  • Delegated: doctrine, task intent, bounds, and authorized evidence let the executor select or construct the choice.
  • Irrelevant: every remaining variation is acceptable and decoupled, so no common choice is needed.
  • Gap: none of those sources closes the choice, so the executor returns or escalates instead of projecting an answer.

A packet need not label every omission. It must carry enough task-specific intent and bounds for consequential cases to remain distinguishable. Intent does not entail privileged facts, permissions, ownership, external commitments, or acceptance evidence; those remain explicit when doctrine cannot determine them. A claimed inheritance is also a gap when the runtime does not actually deliver the doctrine with binding force.

A control regime governs consequential surfaces

A handoff is consequential here when the surrounding system cannot treat a plausible execution-boundary failure as harmless. Examples include a failure that changes acceptance of composed work, violates authority or an external commitment, contaminates peer work, or prevents required recovery. Every surface whose failure could produce such a consequence needs an operative control path.

One non-exhaustive Commonplace audit heuristic checks scoped authority; usable context and resources; output ownership and integration; feedback and verification; accountability and an attributable authority path; and recovery or escalation. Coupled work also needs consistency, common selection, or an actor empowered to resolve incompatible local projections, because agent orchestration needs coordination guarantees. One mechanism may govern several surfaces, and a task may expose additional ones such as timing, liveness, or concurrency.

This audit is a Commonplace inference, not a list taken from military doctrine. The modern sources provide analogues for authority, resources, coordination, reporting, supervision, accountability, and retained responsibility. They do not derive a recovery protocol for agent work. Recovery remains a target-side requirement whose form must be selected from the task's consequences. Prompts, tool policies, sandboxes, workflow roles, and orchestrators can carry controls, but none is sufficient merely by being present.

Outcome wording does not supply the regime

An outcome can name a destination while leaving authority, binding upstream facts, resource access, output ownership, verification, integration, and return of control unsettled. A short message inside a workflow that governs those surfaces can participate in an adequate regime. A detailed outcome description that leaves them unsettled cannot establish one. Structural adequacy therefore turns on the control relation, not prompt length or outcome-only wording. Concision can result when a verified shared path already carries the standing interpretive machinery; it is not evidence that the task relation is adequate. This is not an empirical ranking of prompt styles. More text does not itself supply more control, and this record does not show that a compact methodology cue reliably activates the intended regime.

Scope

This qualitative structural test covers later-informed delegation: the choice is consequential, execution can expose evidence that discriminates among permitted means, the executor is competent and authorized, and the surrounding system governs the failure-relevant surfaces. If all choice-relevant inputs are available and stable upstream, the later-informed rationale does not apply and the choice may be fixed there. Waiting alone creates no information advantage. Leaving a harmless, decoupled choice open also does not show that this mechanism transferred. Other reasons for delegation, including isolation, specialized access, or workload partition, are not derived here.

Hostile purpose, violence, tempo competition, military hierarchy, rank, force structure, professional culture, and military risk doctrine are not requirements for non-adversarial transfer. The target must independently supply the functions it needs: competence, shared meaning, usable information and resources, scoped authority, coordination, feedback, integration, verification, accountability, and recovery. This functional restatement does not claim that a prompt recreates a military institution.

The record does not select a competence test, consequence threshold, or task-specific control implementation. It supplies no effect size, comparative outcome, universal decentralization rule, LLM evaluation, or evidence that a bare historical or methodology label activates the mechanism. An explicit mechanism gloss avoids relying on Auftragstaktik or mission command as an exact specification.

Operationalized into:

  • Writing conventions for kb/instructions/ — makes consequential worker packets task-specific deltas from verified Commonplace doctrine
  • Write an instruction — selects delegation controls from consequential failure surfaces instead of requiring a universal packet template
  • cp-skill-write-multistage — fixes the commission, evidence and authority boundaries, integration, review, and recovery while leaving bounded investigative and authorial means to workers