Task Pilot

Task decomposition and execution framework for Claude — breaks complex work into verified subtasks, plants context anchors to survive compaction, and never l...

Security Scan

An instruction-only skill whose behavior and requirements match its stated purpose: task decomposition, checkpoints, and context anchors; it requests no credentials, installs, or file access.

Purpose & Capability

The name/description match the runtime instructions: the SKILL.md is a protocol for planning, step checkpoints, anchors, recovery, and completion. No unrelated environment variables, binaries, or config paths are requested, and the optional 'pairs well with' install suggestions are reasonable adjuncts.

Instruction Scope

Instructions are narrowly scoped to breaking tasks into plans, planting short context anchors in the conversation, and following a recovery protocol. One operational note: anchoring every 8–12 exchanges means the agent will write repeated structured summaries into chat history — do not include secrets or sensitive tokens in plans/anchors as those will persist.

Install Mechanism

No install spec or code files are present (instruction-only), so nothing is written to disk or downloaded during install. This is the lowest-risk install profile.

Credentials

The skill declares no required environment variables, credentials, or config paths. There are no disproportionate secret or credential requests.

Persistence & Privilege

always:false and default autonomous invocation are used (normal). The skill does not request permanent presence or system-level modifications. As with any agent-invokable skill, consider that the agent can call it autonomously, but that is standard behavior and not by itself a concern here.

Assessment

This skill appears internally consistent and low-risk: it only provides an agent protocol and asks for nothing sensitive. Before using, avoid putting passwords, API keys, or other secrets into task plans or context anchors (those summaries persist in chat). If you install the related ThinkStack skills, review them separately — this evaluation covers only task-pilot itself.

Never lose your place in a complex task. Task Pilot breaks work into verified steps, plants context anchors, and recovers gracefully when conversations get long — so Claude never goes off track or forgets where it was.

The Core Problem

Long tasks break in two ways:

1. Scope drift — execution slowly wanders away from the original goal
2. Context loss — compaction or a long session wipes the working state and Claude starts guessing

Task Pilot solves both.

When to Activate

Use Task Pilot when:

• A task has 3 or more distinct steps
• Execution will span many exchanges
• An error in one step would cascade to the next
• You need to hand back to the user at milestones

Skip it for: single-step tasks, quick lookups, tasks that take one response.

The Protocol

Before Starting: Plan First

Before writing a single line of code or taking any action, create a task plan:

[Task Plan]
Goal: ...
Steps:
  1. [ ] ...
  2. [ ] ...
  3. [ ] ...
Success criteria: ...
Risks / dependencies: ...

Wait for confirmation (or adjustment) before executing. Never skip the plan.

During Execution: Step Checkpoints

After completing each step:

• Mark it done: [✓ Step N — description]
• State what was produced
• State what comes next
• Flag any deviation from the plan and explain why

If a step fails, say so immediately. Do not silently move to the next step.

Context Anchors

Every 8–12 exchanges, or before any major transition, plant a context anchor:

[Context Anchor]
Task: ...
Completed: steps 1–N ✓
Current: step N+1 — [description]
Remaining: steps N+2 to end
Key decisions locked: ...
Blockers / open questions: ...

Anchors are short. They exist to let you — or a future session — resume from exactly this point.

Recovery Protocol

If context appears lost or compacted:

1. Find the most recent Context Anchor
2. Read the task plan to identify completed steps
3. Resume from the last incomplete step
4. Never redo completed work unless the user explicitly asks

When recovering, announce it:

[Recovering from context loss]
Last anchor: step N complete
Resuming at: step N+1

Completion Gate

A task is done when:

• All steps are marked ✓
• Success criteria are met
• Side effects (files changed, state modified, services restarted) are verified
• A brief completion summary is given

[Task Complete]
Completed: all N steps ✓
Result: ...
Verified: ...

Scope Discipline

• Do not expand scope without re-running the plan step
• If the user adds requirements mid-task, pause, update the plan, get confirmation
• "While I'm here I'll also..." is a scope violation unless explicitly approved

Pairs Well With

• clarity-first — clarify requirements before creating the plan
• thinkdeep — analyze complex decisions within a step before executing

Install the full ThinkStack for best results:

openclaw install clarity-first
openclaw install thinkdeep
openclaw install task-pilot