Use this skill when you need to finish writing a block of code that relies on functions or classes defined elsewhere in the same project rather than standard libraries. It should be triggered by layman requests such as 'finish this function using our internal helpers', 'complete this call but I don't remember the exact sub-module function', 'help me use our custom API to register this dataset', or 'fix my code where I'm guessing the name of our team's functions'. Do not use it for general Python library completion; use it for project-specific internal logic that isn't found in online docs.
[Case 1]
register_cityscapes.py contains a logic block for registering semantic segmentation metadata but is missing the final internal function call to load the data.DatasetCatalog.register(sem_key, lambda x=image_dir, y=gt_dir: [code to complete])load_cityscapes_sem_seg. Using this identifier as a seed, the agent queries the internal API knowledge base and discovers a functionally similar, actual function called load_cityscapes_semantic in datasets/cityscapes.py. The agent then verifies the parameters of the real function and updates the code to use the correct name and argument structure.[Case 2]
DefaultTrainer class in trainers/base.py and needs to initialize the logger writers using project-specific utilities.self.log_writers variable within the __init__ function of the DefaultTrainer class using the appropriate project-defined utility.getattr pattern to find common writers. Realizing this pattern is likely a placeholder, the agent summarizes the intent as 'obtaining log writers from project settings' and searches the functional semantic index. It identifies the internal API get_log_writers in the project's utility module, which perfectly matches the required behavior. The agent then implements the call while passing the existing self.opts and self.save_location variables.To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:
Phase 1: Environment Exploration
Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.md
Phase 2: Trajectory Selection
Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.md
Phase 3: Data Synthesis
Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md