# Attention Management > Restore and maintain the governing relation for an Agent's next action when a long task drifts into a local branch, a correction changes assumptions, a phase transition buries the mainline, or recurring reminders, role-entry, or prompt habits need evaluation. Use for "manage agent attention", "return to the mainline", "attention drift", "保持主线", "注意力管理", "提示词衰退", or when the governing relation has drifted and a branch needs retention, return, reset, or reform judgment. Do not use to author doctrine, plan an entire project, configure vendor hooks, or infer neural activation from behavioral output. - Skill: `lidessen/attention-management` (Agent Skill, multi-file: 6 files) - Install (CLI): `npx skillmds@latest add lidessen/attention-management` - Raw SKILL.md: https://api.skillmd.com/api/skills/lidessen/attention-management/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: lidessen (https://skillmd.com/u/lidessen) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lidessen/attention-management --- # Attention Management ## Principle expression **Primary:** P09 **Supporting:** P04, P03, P15 ## Scope Own one recurring judgment: **at an Agent decision boundary, which governing relation should occupy active attention next, and what is the smallest reorientation that restores it?** This Skill manages the live relation among task, authority, acceptance, evidence, and current action. It does not own their source meaning. It does not turn a todo list into a semantic authority, make every local issue part of the mainline, or preserve an obsolete goal after the Principal has switched tasks. Use `context-engineering` when the unresolved judgment is which runtime surface can deliver source information. Use `agent-delegation` when the relation is settled and the question is which bounded contribution should leave Main. Use `practice-cycle` after an observation when the unresolved judgment is the next practice rather than the active governing relation. ## Principle source Use the host Sequence and the matching P09, P04, P03, and P15 interpretations when the host declares them. Otherwise use the read-only fallback in `references/sequence.md`. ## Dispatch - For a live drift, correction, or branch-ownership judgment, use the core method below and load no optional reference by default. - When the failure is difficult to distinguish from context delivery, memory, planning, or task switching, read `references/concepts.md`. - When constructing a working-character, role-entry, catchphrase, or habit candidate, read `references/disposition-carriers.md`. - When comparing any attention treatment, read `references/evaluation.md` and load the carrier reference only if that treatment's semantic atoms must be inspected. ## Start Recover only the state that can change the next action: ```text Authoritative active task and current acceptance: Current phase and next consequential decision: Governing relation that should shape it: New observation, correction, or distracting branch: Why the branch may or may not affect acceptance: Current candidate action: Observation that would show the attention judgment is wrong: ``` If the task has already completed, paused, been handed off, or been replaced by a later Principal message, form the new task. Do not call that a return to the old mainline. ## Core method 1. **Form the current object.** Read the active human message and its relevant source state. Separate the task that is authoritative now from an older goal, an implementation plan, and the loudest local problem. Do not infer the object from a preferred Principle or slogan. 2. **State the execution anchor.** Compress only the active goal, acceptance, authority/source boundary, current phase, governing relation, and retained obligations. This projection is replaceable and may be rebuilt from sources; it is not a second task record. 3. **Classify the disturbance.** Choose one: - `continue`: the current action still advances the anchor; - `retain`: the branch blocks acceptance, invalidates a load-bearing assumption, or changes the governing relation, so it belongs in the active line for now; - `return`: the branch is settled, non-blocking, or evidence-only, so restore the next anchor action; - `switch`: the Principal completed, paused, handed off, or replaced the active task, so reconstruct a new anchor; - `ask`: authority or meaning is genuinely ambiguous and different answers would materially change the transition. 4. **Choose the smallest control.** Take no control when attention is already correct. Use a compact cue when the anchor is valid but buried. Use a reset when a correction or evidence rejects an assumption: replace the obsolete attention unit with the new invariant instead of appending another reminder. Re-form the anchor on a true task switch. Keep a load-bearing branch active only until its named return condition. 5. **Rebind an existing candidate action—or hand off.** When one candidate action already exists, verify that it changes acceptance, resolves the retained relation, or obtains evidence that decides between them. When the governing relation is restored but the next practice remains undecided, return the reconstructed anchor and observations to `practice-cycle`; do not select the practice here. Externalize secondary obligations rather than recursively following every new defect. 6. **Observe the transition.** Check the first action after a boundary, hard constraints, irrelevant branch depth, return latency, and whole-task acceptance. Vocabulary, confidence, role declaration, or reminder reproduction do not prove correct attention. 7. **Cultivate only from recurrence.** When the same failure repeats, inspect only the matching optional reference from the dispatch above. Preserve hypotheses and negative results; do not silently promote them into the core method. ## Compact result ```text Execution anchor: Disturbance: continue | retain | return | switch | ask Reason it changes or does not change acceptance: Control: none | cue | reset | reform | clarify Rebound candidate action or handoff: Retained obligation and return/reopen signal: Evidence to observe: ``` Do not emit this packet merely to narrate ordinary progress. Use it when it changes the next action, restores a lost mainline, or creates reconstructable evidence for an attention treatment. ## Boundaries - Do not claim that model weights changed, neural activity persisted, or Transformer attention decayed from prompt behavior alone. - Do not repeat a full system prompt merely because a task is long. Selective reactivation is a treatment to test, not a ritual. - Do not let a catchphrase, character, plan, summary, or status field outrank the Principal message and authoritative sources it projects. - When a delegated task needs owner/status/dependency visibility, use the existing coordination projection owned by the development method; this Skill restores attention to the governing relation but does not become the project plan or delegation tracker. - Do not force every task through an orientation ceremony. Additional reasoning, latency, and context are costs and possible confounders. - Do not use a return-to-mainline cue after a genuine task switch. - Do not route through Workbench, Work Cell, or a vendor hook unless their separately owned persistent state or runtime capability is actually needed. ## Completion standard Attention is restored when the current task and governing relation are source linked, the disturbance has a justified disposition, the next action advances acceptance or resolves a load-bearing branch, obsolete anchors no longer govern the task, and any claimed treatment effect has matched behavioral evidence.