About this skill

Problem

Long-running agent tasks can lose focus when context is treated as ephemeral. Multimodal information may disappear after inspection, browser screenshots are not persistent, and stage state, decisions, and errors are often left only in memory. When the context window refreshes, the agent may repeat work or drift from the original goal.

How It Works

This skill makes planning explicit by writing durable files into the project directory. It uses task_plan.md for stages, progress, and decisions; findings.md for research conclusions and key discoveries; and progress.md for session logs, test results, and operation traces. Before starting, create a plan from the templates; while working, update stage status, record errors, and reread the plan before major decisions to keep the goal active in the context window.

It emphasizes converting multimodal observations into text. After viewing images, PDFs, or browser data, important findings should be saved to findings.md instead of relying on screenshots or context memory. When resuming a session, the agent can reorient around five questions: current stage, remaining stages, goal, what has been learned, and what has been done.

Boundaries

Use it for multi-step work, research tasks, build projects, and workflows that span many tool calls. Skip it for simple questions, single-file edits, or one-shot answers. The planning files should live in the current project root, not in the skill installation directory. If a repository already has mature task-tracking or documentation conventions, treat this skill as the agent-facing memory layer rather than a replacement for repo docs.

Screenshots

Use Cases

When researching code or web pages, save key findings to findings.md before screenshots or multimodal context is lost.
For build tasks spanning multiple tool calls, mark stage status in task_plan.md and log test results in progress.md.
After resuming an interrupted session, reread task_plan, findings, and progress to confirm the current stage and goal.
When debugging repeated failures, record each attempt and error in task_plan.md to avoid silently retrying the same path.

Best For

Engineers running multi-step development tasks with an agent, who need stage state, decisions, and errors persisted to files.
Researchers or docs maintainers performing technical research, who need web pages, PDFs, and code findings saved durably.
Claude agent users executing complex tasks, who need to recover stage and attempted approaches after a session break.
Developers maintaining multi-file projects, who need errors, test results, and stage status written into the repo, not context.