implement-experiment

Implement Experiment Skill

Implement an experiment plan in an isolated git worktree. All experiment artifacts are created inside a single self-contained folder under research/. The worktree is left intact for the orchestrator to run the experiment, test, and merge separately.

This skill reads the experiment plan and follows its implementation phases. The plan specifies the directory layout, what scripts to write, what data to generate, and what environment to set up. This skill builds all of it.

When to Use

• As the implementation step of the research recipe (phase 2)
• After the experiment plan has been approved via GitHub issue

Arguments

/autoskillit:implement-experiment {plan_path}

{plan_path} — Absolute path to the experiment plan file (required). Scan tokens after the skill name for the first path-like token (starts with /, ./, or .autoskillit/).

Critical Constraints

NEVER:

• Implement without first exploring affected systems with subagents
• Implement in the main working directory (always use the worktree)
• Force push or perform destructive git operations
• Merge the worktree branch into any branch
• Delete or remove the worktree
• Run the full test suite — pytest with no args or targeting the entire repo (the orchestrator handles full test execution via test_check)
• Create experiment files outside the planned research/ subfolder
• Execute git merge commands (all branch content must be applied via git cherry-pick or git checkout <branch> -- <file>)
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Create a new worktree from the current branch
• Use subagents to deeply understand the codebase context BEFORE implementing
• Use model: "sonnet" when spawning all subagents via the Task tool
• Follow the implementation phases from the experiment plan
• Put all experiment artifacts in one self-contained research/ subfolder
• Commit per phase with descriptive messages
• Leave the worktree intact when done
• Write tests/test_{script_name}.py alongside each experiment script created in Step 4
• Run pytest --collect-only after creating tests to verify discovery before committing
• Read before editing: Before issuing an Edit call on any file, ensure you have issued a Read on that file earlier in this session. Claude Code rejects Edit on unread files — the retry wastes a full API turn at current context size. If you are uncertain whether a file was read, issue a targeted Read (offset + limit to the region you plan to edit) rather than risk an error.