base/.agents/skills/execute-tasks/SKILL.md

name, description

name	description
execute-tasks	Execute and complete a specified list of project tasks sequentially. Use when asked to work on, implement, execute, or do a set of task IDs. Supports writing, research, planning, software, and other projects; verifies results, records human review steps, and transitions completed work to testing.

Execute Tasks

Orient

1. Read AGENTS.md, project-config.json, and the task tracker.
2. Load the task-management skill and query each requested task with ./scripts/get-task <ID>. Do not assume a prefix or software project.
3. Confirm tasks are ready for work. If not, explain the status and ask for direction; do not silently bypass blocked or unscoped work.
4. Inspect existing materials and changes. Preserve unrelated work.

Execute Sequentially

For each task:

1. Transition it to in_progress using update-task.
2. Resolve material ambiguities with the user and agree on acceptance criteria.
3. Complete the requested deliverable using the project's established methods. For a novel this may be manuscript edits; for research it may be source analysis; for software it may be code. Do not invent a language or platform.
4. Update directly related documentation and verify acceptance criteria. Use configured commands only when applicable. For non-command workflows, review deliverables and report what still needs human judgment.
5. When editing the bundled task tooling, run python3 -m unittest discover -s scripts/tests -v.
6. Record step-by-step human verification in docs/task-verification.md, creating it if needed. Include expected results and limitations.
7. Transition successful work to testing with --check-all, then query the task again to verify the transition. Never mark your own work done.
8. Record newly discovered work with add-task; do not expand scope silently.

Stop and explain blockers when an acceptance criterion cannot be met. Do not present unverified work as complete.

Persistence

• Do not commit unless asked. If asked, inspect git status and include the task/verification changes relevant to the deliverable. Do not stage unrelated edits blindly.
• Use descriptive commit messages including task IDs and the required co-author trailer where applicable.
• If project-description documents exist, check their last update and refresh verification timestamps after major changes or 20 commits. Do not require nonexistent project-specific documents.
• Human reviewers move testing to done and migrate verification steps to docs/human-tests.md, creating it when first needed.
• If CLI tools cannot run, disclose that and agree on a manual workflow with the user rather than reporting fabricated checks.