Clear Task
/install cleartask
Clear Task
Say it clear, do it right.
Purpose
Use this skill to turn an underspecified user request into an internal AgentTaskGuidancePackage, then execute the user's task using that package. Do not display the full Stage3 package unless the user explicitly asks to inspect, export, save, or review it.
Workflow
Run Stage1 and Stage2 first. After Stage2 produces the UniversalRequirementHandoff, stop and ask the user to confirm before Stage3. Do not run Stage3 or execute work until the user explicitly approves the Stage2 handoff.
- Stage1: Build a rich
UniversalProcessModel. - Stage2: Clarify missing details and produce
UniversalRequirementHandoff. - Confirmation Gate: show a concise user-facing requirement summary and ask the user to confirm, revise the requirement description, or revise requirement details.
- Stage3: Generate an internal
AgentTaskGuidancePackageusing expert-role task guidance and expert-role acceptance criteria. - Execute the task using the internal package.
- Report the result, key decisions, created/changed files, and concise self-check.
Do not add a Stage0 intake gate. If the request is invalid, unsafe, or impossible to model, mark it as out of scope inside Stage1/Stage2 and explain the blocker.
Stage1
Model the user's request as a workflow or work process, even when the request is a product, document, game, analysis, or implementation task. Prefer a graph over a flat step list.
Read references/stage1-universal-process-model.md when you need the full schema or need to output a Stage1 artifact.
Minimum Stage1 output:
request_summarybusiness_goalscopeactorsinputsoutputsworkflow_graph.nodesworkflow_graph.edgesbusiness_rulesassumptionsopen_questionssource_evidence
Keep unconfirmed model inferences in assumptions or open_questions; do not present them as confirmed facts.
Stage2
Use Stage2 to close the gaps that would materially change downstream work. Ask the fewest high-value questions needed. When the user asks for recommendations, choose reasonable defaults and record them under accepted_defaults.
Read references/stage2-clarification-handoff.md when generating clarification questions or the final Stage2 handoff.
Stage2 is ready for Stage3 when:
- blocking questions are answered, or
- blocking questions have accepted defaults, and
- unresolved details are non-blocking or explicitly left to the downstream agent.
Stage2 being ready does not authorize Stage3. It only means the requirement handoff is ready for user confirmation.
Confirmation Gate
After Stage2, always pause and ask for explicit user confirmation before Stage3. Do not continue from Stage2 to Stage3 in the same response.
Show a concise user-facing confirmation summary using plain requirement language, not internal stage terminology:
需求描述: one short paragraph describing what the user wants.需求细节: scope, deliverables, constraints, accepted defaults, assumptions, and non-blocking open questions.需要你确认: ask whether to continue.
Offer these user-facing choices:
确认,继续修改需求描述: ...修改需求细节: ...
Treat only explicit approval as confirmation, such as "确认", "继续", "可以", "批准", "go ahead", or an equivalent clear instruction. If the reply is ambiguous, ask a short confirmation question and continue to pause.
If the user asks to modify the requirement description:
- Update the Stage1
UniversalProcessModel. - Re-check whether Stage2 changed because of the revised description.
- Revise or regenerate the Stage2
UniversalRequirementHandoff. - Return to this confirmation gate.
If the user asks to modify requirement details:
- Update the Stage2 clarification/handoff content, including scope, defaults, assumptions, constraints, deliverables, and open questions as needed.
- Do not run Stage3 automatically.
- Return to this confirmation gate.
If a revision introduces new blocking questions, ask only those questions, update Stage2 from the answers, then return to this confirmation gate.
Stage3
Do not start Stage3 unless the user has explicitly confirmed the Stage2 handoff through the Confirmation Gate.
Stage3 always produces an AgentTaskGuidancePackage, but this package is internal execution context by default. It does not choose whether the final downstream result should be code, PRD, PPT, Word, test plan, or another artifact. The executing agent decides that from the task package and continues the work.
Stage3 must use expert roles:
- Classify the task domain and downstream work mode.
- Select one primary expert and supporting experts.
- Generate expert task guidance.
- Generate expert acceptance criteria.
- Compose a consolidated package without inventing new requirements.
Read:
references/stage3-agent-task-guidance-package.mdfor the internal package schema and execution policy.references/expert-roles-and-acceptance.mdfor expert selection and acceptance criteria rules.
Execution Policy
After Stage3, continue execution when:
- the user has explicitly confirmed the Stage2 handoff,
- blocking questions are resolved or have accepted defaults,
- the user asked for a concrete deliverable or outcome,
- the task can be completed with available tools and workspace access.
Pause and ask before execution when:
- the user has not explicitly confirmed the Stage2 handoff,
- Stage2 has unresolved blocking questions,
- execution would unexpectedly modify many files or external systems,
- the task requires credentials, paid services, network actions, or destructive operations,
- the user explicitly asked only for analysis, planning, or the internal handoff package.
Output Rules
- Preserve traceability from user input, Stage1, Stage2, and expert recommendations.
- Separate
confirmed_facts,accepted_defaults,assumptions, andexpert_recommendations. - Use blocking questions only for decisions that materially alter downstream work.
- Acceptance criteria must be verifiable and include pass conditions.
- Avoid vague criteria such as "good UX", "complete feature", or "high quality" unless converted into observable checks.
- If expert guidance conflicts, keep the conflict in the package and suggest a conservative resolution.
- Do not display the full
AgentTaskGuidancePackageby default. - Use expert acceptance criteria for execution self-check and final concise reporting.
Internal Package Shape
Use this shape internally. Do not render it to the user unless explicitly requested.
AgentTaskGuidancePackage:
title:
objective:
expected_outcome:
downstream_agent_role:
confirmed_scope:
requirements:
process_model:
expert_task_guidance:
consolidated_execution_plan:
expert_acceptance_criteria:
consolidated_acceptance_criteria:
risks_and_edge_cases:
unresolved_questions:
assumptions:
conflicts:
traceability:
suggested_next_prompt:
Default User-Visible Result
Before Stage3, report only:
- requirement description,
- requirement details,
- accepted defaults and assumptions,
- remaining non-blocking questions,
- a clear request to confirm, modify the requirement description, or modify requirement details.
After execution, report only:
- what was completed,
- key decisions or defaults used,
- created or changed files/artifacts,
- self-check against must-pass criteria,
- remaining caveats or non-blocking questions.
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install cleartask - After installation, invoke the skill by name or use
/cleartask - Provide required inputs per the skill's parameter spec and get structured output
What is Clear Task?
Transform broad, ambiguous, or early-stage user requirements into executable work through Stage1 requirement understanding, Stage2 clarification, a required... It is an AI Agent Skill for Claude Code / OpenClaw, with 84 downloads so far.
How do I install Clear Task?
Run "/install cleartask" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Clear Task free?
Yes, Clear Task is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does Clear Task support?
Clear Task is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Clear Task?
It is built and maintained by Hengjing Zhang (@zhanghengjing2); the current version is v1.0.0.