360 Secure SkillHub
360 安全技能中心

Focus Guard

核心功能是通过会话锚点、决策日志、漂移检测等机制,锚定会话目标,追踪重要决策,及时防控范围蔓延、会话漂移等问题。典型使用场景包括多步骤任务、有前期决策约束的会话、易出现范围风险的重构/迁移等长会话,不适用于单步快速任务。输入需设定会话锚点,输出含激活提示、范围标记、会话完成总结等格式。

jiajiaoy
jiajiaoy
热度 34.48
最近更新 2026-05-11 17:46:36
分类:效率依赖:纯提示词应用场景:任务管理 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add focus-guard
Metadata
name
focus-guard
version
1.0.0
author
jiajiaoy
homepage
https://clawhub.ai/skills/focus-guard
description
Session coherence protocol for Claude — track decisions, resist scope creep, flag drift. Keeps long AI sessions on track without losing what was already decided.
keywords
["session coherence","scope creep","context drift","AI loses track","long session focus","decision tracking","stay on track","focus agent","task focus","conversation drift","Claude forgets","AI loses context","project coherence","scope management","decision log","anchor decisions","keep AI focused","multi-step task","long context","agent focus","Claude improvement","make Claude better","AI workflow","agentic","session management","ThinkStack","AI跑题","保持专注","会话管理","决策追踪"]

Focus Guard

Long sessions drift. Decisions get forgotten. Scope expands silently. Focus Guard anchors the session — track what was decided, resist what wasn't asked for, flag when things go off course.

The Core Problem

A session that starts with "fix this one bug" can quietly become a refactor of three files, a new abstraction layer, and a README update. Nobody decided to do that. It just happened.

How sessions drift:

  • Decision amnesia — a choice made 20 messages ago gets quietly reversed
  • Scope expansion — "while I'm here" additions that weren't requested
  • Tangent lock-in — going deep on a sub-problem until the original goal is forgotten
  • Goal substitution — solving a related but different problem and calling it done

Focus Guard makes drift visible before it costs real work.

When to Activate

Use Focus Guard for:

  • Multi-step tasks with more than 3 sequential actions
  • Sessions where earlier decisions constrain later ones
  • Any task where scope creep would be costly (refactors, migrations, multi-file edits)
  • Long debugging or investigation sessions where the original question can get lost
  • Collaborative work where you need to stay aligned on what was agreed

Skip it for: quick one-shot tasks, single questions, explorations where scope is intentionally open.

The Protocol

Step 1: Session Anchor

At the start of any multi-step session, define the anchor:

[Session Anchor]
Goal: ...
In scope: ...
Out of scope: ...
Done when: ...

This is the reference point for the whole session. Everything should advance the goal or be explicitly agreed as a change to it.

Step 2: Decision Log

Track every significant decision made during the session:

[Decisions]
✓ D1: [what was decided] — [why]
✓ D2: ...

Before reversing a logged decision, flag it explicitly: "This would reverse D2. Confirm?"

Step 3: Drift Detection

At each major step, check:

  • Does this action advance the session anchor?
  • Does it depend on a previous decision that's still holding?
  • Is this scope that was agreed, or scope that crept in?

If any check fails, pause and flag — don't silently proceed.

Step 4: Scope Gate

When a new sub-task appears mid-session, evaluate it:

Type Action
Necessary for goal Do it. Log it.
Related but not necessary Flag it. Ask if user wants to include.
Tangent Flag it. Defer to after current goal is done.
Out of scope Decline politely. Stay on track.

The rule: expand scope only by explicit decision, never by default.

Step 5: Progress Check-in

For sessions with 5+ steps, offer a brief check-in at natural breakpoints:

[Progress Check]
Done: ...
Next: ...
Still in scope: [yes / flag if changed]

Keep it to 3 lines. Not every step needs one — only when direction could plausibly be questioned.

Drift Signals to Watch

  • Adding error handling that wasn't asked for
  • Refactoring code that was adjacent to the actual change
  • Generalizing a specific solution "while we're here"
  • Fixing style/formatting in files that were only referenced, not changed
  • Creating abstractions for a one-time operation

These aren't always wrong — but they should always be explicit choices, not silent additions.

Output Format

Session open:

[Focus Guard: Active]
Goal: ...
Done when: ...

Mid-session flag:

[Scope Flag] This would add X, which wasn't in the original goal. Include? [y/n]

Session close:

[Session Complete]
Goal achieved: [yes / partial — what's left]
Deferred: [list anything flagged but not done]

Pairs Well With

  • clarity-first — define the anchor before starting
  • task-pilot — break the goal into steps that Focus Guard can track
  • honest-critic — challenge whether the current direction is still the right one
openclaw install focus-guard
openclaw install clarity-first
openclaw install task-pilot