learn from my corrections

The problem

Users waste time repeating the same corrections, stylistic preferences, and project requirements across multiple chat sessions. AI assistants often fail to internalize specific feedback, leading to a cycle of repetitive editing and frustration.

What it does

• Analyzes conversation history and rejected drafts to identify recurring patterns in user feedback.
• Extracts concrete, observable instructions from explicit corrections and positive examples.
• Classifies findings into global preferences, project-specific rules, or one-off factual fixes.
• Generates a ready-to-use instruction block with evidence-based confidence ratings for each rule.
• Maps instructions to their ideal destination, such as system prompts, project files, or custom skills.

Why this beats prompting it yourself

Manually distilling feedback often leads to over-generalization or "prompt bloat" where conflicting instructions confuse the model. This tool uses a structured evidence ledger to ensure new rules don't break existing workflows or ignore legitimate exceptions. It prevents the common mistake of turning a single correction into a blanket ban.

Use cases

• Consolidating feedback from a long project thread into a permanent project README or system prompt.
• Converting a series of stylistic critiques on marketing copy into a reusable brand voice guide.
• Auditng why an assistant keeps failing a specific task by comparing corrections against current instructions.

Known limitations

Cannot access history from other conversations unless the user provides the transcripts or the environment explicitly supports cross-thread retrieval. Does not "train" the underlying model.