Attention Driven

⚡ Main-Agent Protocol — 主 Agent 操作协议（压力下先读这 ~60 行）

This SKILL.md is for the Main Agent only. Do not pass it wholesale to a worker. The delegation contract lives in references/delegation.md; worker execution rules live in references/worker.md.

MVP: Three Rules

Rule 1: Name the 30% Before Acting

任何改变状态的工具调用前，输出一行： 30%: <当前主要矛盾，一句话> 说不出 → 不要动手。

Rule 2: Delegate the 70%

Main Agent 只守 30%：主要矛盾、约束、路由、验收标准、方向修正。

任何会展开成 70% 执行细节的工作都要委托：实现、广泛搜索、调试、 局部验证、逐文件检查、重复工具循环。

委托时只给 worker delegation envelope 和 references/worker.md 里的执行 规则。envelope 模板在 references/delegation.md。不要把 Main Agent 的 完整 skill 交给 worker。

Rule 3: Verify the Compressed Return

Main Agent 只消费 worker 的压缩返回：结论、证据、风险、变更路径、未决点。 不要把完整执行轨迹吸进主上下文，除非证据不足或 30% 本身存疑。

worker 返回后，对照验收标准。结果不符 → 不重复同一委托。修正 delegation envelope，或回到规则一重新审视 30%。

同一委托因同一原因失败两次 → 假设被证伪。执行识别：

假设: <哪个假设被证伪了？>
新方向: <基于被证伪的假设，30% 应该调整为什么？>

🚧 Delegation Envelope（进入 70% 前，不可跳过）

当下一步会展开成 70% 执行细节时，读取 references/delegation.md，先输出 其中定义的 envelope，再委托给 worker。

Main Agent 可以直接做 30% 工作：读取入口状态、判断 route、修正目标/设计/ 验收标准、审阅压缩返回、记录 durable artifact。不要下场做会污染主上下文的 70% 执行。

🗺️ 层面路由（强制步骤）

进入新层面之前输出两行。禁止在输出前调工具：

当前层: [goal/design/fact/reframe/harness]
原因: <为什么是这一层，一句话>

Attention-Driven

attention-driven is the stable entrypoint for agent work. It keeps one mainline through durable artifacts instead of making the agent choose between competing methodology skills.

The entrypoint is intentionally Main-Agent-only. Keep worker execution policy out of this file; load references/delegation.md and references/worker.md only when constructing a bounded delegation prompt.

Core Law

Hold the load-bearing 30% and give the remaining 70% freedom. Spend most attention on the slow-changing skeleton: the goal line, design shape, fact claim, or reframe primitive that determines the rest. Treat the surrounding execution detail as flexible and cheap to retry.

The 30% earns that budget for two reasons. Attention is finite, and the skeleton determines the rest — mistakes there make every downstream choice more expensive. And the 30% is the structural layer: costly to change once committed, while the 70% stays cheap throughout. A renovation makes the asymmetry concrete. Soft furnishing drives most of the visible result and is cheap to swap; plumbing, wiring, and load-bearing walls gate what can happen later. A good shape leaves few load-bearing walls in the way — merging two rooms is just taking a wall down. A bad shape made every wall load-bearing by accident, and now nothing can move.

The corollary is design intent. Put something in the 30% only when it needs to be load-bearing, and keep the 70% structurally cheap to change. When a 70% item has quietly become load-bearing — callers everywhere depend on its shape, or removing it would cascade — it has migrated into the 30% without a decision. Catch that before the next change pays for it: either ratify the new load and route it through design, or restore the boundary so the 70% remains free.

Apply the same rule to decisions. Maximize system agency for reversible and local choices: decide, delegate, review when useful, verify the claim, and move on. Escalate only the load-bearing 30% to the human: goal changes, irreversible tradeoffs, authority gaps, value judgments, or decisions whose failure would invalidate the rest.

Autonomy is the default inside accepted boundaries. If the next move is reversible, observable, within authority, and aligned with the current goal/design, the Main Agent decides and steers: preserve the 30%, choose the route, write the envelope, verify the compressed return, or update the durable artifact. Use reviewer attention for uncertain 70% choices. Use human attention only when the decision changes direction, authority, values, irreversible cost, or durable system shape.

Model work as something that can be steered: name the target, understand the current shape, observe the gap, make the smallest useful correction, and look again. If the gap does not shrink, improve the observation, change the correction, reject the disturbance, or update the understanding instead of pushing harder.

Keep loop speeds separate. Goals and harness setup are slow; design is medium; fact checks and execution are fast. Fast-loop noise should not rewrite slow-changing artifacts until repeated observations show the model is wrong.

Read the layer names as local steering responsibilities, not as phases:

• goal keeps direction and success criteria.
• design keeps the current system shape and boundaries.
• fact keeps observations that can prove or falsify claims.
• reframe changes the lens when the old category cannot guide the work.
• harness keeps project context wired so future agents can continue. Includes the HANDOFF.md convention — an ephemeral state snapshot written by the outgoing agent, read first by the incoming agent, consumed and archived after a single use.

The same rule applies when maintaining attention-driven instructions. Command names, artifact wording, and setup mechanics may improve, but those execution improvements must not disturb the core method: find the principal 30%, preserve it durably, and let the agent move freely inside it.

When learning from an external theory or method, absorb its operating principle before borrowing its form. Localize it to the concrete problem, keep what improves steering, and discard ceremony that does not reduce the current gap.

Commands

When invoked with an argument, dispatch to the corresponding command file:

• /attention-driven go -> read commands/go.md. Daily entrypoint: resume context, identify the mainline, route to the smallest needed layer, and close with a record draft when goal artifacts exist.
• /attention-driven setup -> read commands/setup.md. Install or sync the project harness instructions and managed blocks.
• /attention-driven audit -> read commands/audit.md. Check drift across goal, design, fact, reframe, and harness artifacts.
• /attention-driven goal ... -> read commands/goal.md. Set, review, or maintain GOAL.md, records, STOPs, and stories.
• /attention-driven design ... -> read commands/design.md. Work on system shape, decisions, design bootstrap, and blueprints.
• /attention-driven fact ... -> read commands/fact.md. Add falsifiable verification discipline to execution work.
• /attention-driven reframe ... -> read commands/reframe.md. Redraw shape when the paradigm is unsettled.
• /attention-driven close -> read commands/close.md. Close a goal initiative or reframe concept.
• No argument -> behave like go.

Layer Vocabulary

Use these names consistently:

• goal: why, how far, criteria, records, STOPs, principal tension.
• design: current system shape, boundaries, decisions, blueprints.
• fact: observations that can prove or falsify a progress claim.
• reframe: pre-design exploration when the category has no stable shape.
• harness: the project context, instructions, hooks, and artifact wiring that let future agents resume without rediscovery.

Routing Rule

First ask: "What is the principal 30% right now?"

• Direction or success is unclear -> goal.
• System shape or module boundaries change -> design.
• A progress/completion claim needs proof -> fact.
• The shape cannot be stated because the paradigm is unsettled -> reframe.
• Agent context or project wiring is missing/stale -> harness via setup or audit.

Read references/routing.md for the full routing table when the task is ambiguous.

Artifact Rule

Preserve only the load-bearing 30% as durable artifacts: current goal plus recorded output, design plus blueprint/execution direction, fact evidence, reframe primitive, and setup wiring. Leave the 70% strategy flexible and mostly fire-and-forget.

Do not ask the human to decide every local tradeoff. Use agent judgment and agent review for ordinary 70% decisions; surface the few decisions that change direction, authority, values, or system shape.

If an older decision, blueprint, record entry, or concept note was wrong, do not repair history for its own sake. Correct the current artifact or create the next one that supersedes it. Read references/artifact-policy.md before editing durable workflow artifacts.

Progressive Loading

Keep this file as the entrypoint. Load command files for actions and reference files only when the command needs the detail:

• references/goal.md for goal/record/STOP/story semantics.
• references/design.md for 30/70 design, decisions, blueprints.
• references/decision.md for autonomous action, reviewer routing, and human escalation gates.
• references/writing-guide.md for design artifact prose.
• references/fact.md for falsifiability, TDD, evidence trails.
• references/reframe.md for reframe phases and closure.
• references/harness.md and references/setup.md for project wiring.
• references/templates.md for artifact templates.
• references/control-loop.md for the localized engineering-control lens: target, shape, observation, gap, correction, disturbance, stability, and adaptation.
• references/control-review.md for three-role dialectical review in research-like design/proposal work or method imports into a project.
• references/delegation.md for the role visibility matrix, delegation envelope template, context firewall, and worker escalation signal.
• references/worker.md for worker-facing execution rules used inside a delegation envelope. Do not load it as the Main Agent's own operating rule.
• references/effort.md for the agent-effort model: why time is a broken metric for Agent work, and how to use context tokens + tool calls instead.
• references/migration-audit.md for temporary notes recovered from the old split skills that have not yet earned a permanent home.