write-notes-like-deepseek
A Claude Code skill that records the reasoning behind non-trivial code decisions in the repository itself, so later AI sessions see why past choices were made instead of repeating or breaking them.
write-notes-like-deepseek packages a note-keeping method into an installable skill for AI coding agents. It is based on the practice of recording decisions, trade-offs, and verification directly in the repository, so a new AI session reads the reasoning behind past choices instead of guessing.
Before making a non-trivial change, the agent writes a proposal describing the problem and the alternatives it considered, including the strongest case for each rejected option. After the change ships, the note is rewritten in the present tense as a record of what happened. When a decision is superseded, the old note is archived rather than edited: each archived note is sealed with a hash recorded in a manifest, and a script finds and fixes any remaining links to it before archiving is complete.
The rules apply to changes that affect behavior, architecture, cross-file contracts, tooling, test strategy, or on-disk or network formats; purely mechanical changes such as formatting or dependency bumps do not need a note.