discord-manager: Skill for Claude Code

Self-Improvement

Everything here documents the repository as it is on main. If main disagrees with this file, main wins: follow it and flag the drift.

The last step of every task: mine what just happened for durable lessons and codify the ones that earn it. This is the most valuable and most dangerous part of the process — a good instruction improves every future session; a bad one biases every future session. Generalize with the big picture in mind, and codify only with conviction.

When to run

• After every other Definition of Done criterion has passed, before declaring the task done.
• For a multi-task effort, one pass over the whole arc when the effort concludes, with the full picture in view. Individual tasks inside such an effort skip the per-task pass — their lessons are mined at the end, never piecemeal.
• "No lessons worth codifying" is a valid and common outcome, especially for small tasks — state it explicitly and finish. Never manufacture a lesson to look productive.

Source ranking

Weigh evidence in this order:

1. User instructions, corrections, and feedback — the highest signal, outranking everything else. Anything the user had to say twice is something the system failed to absorb the first time: a top-priority candidate. A lesson that contradicts a stated user preference is wrong, no matter how much other evidence supports it.
2. Failures with a root cause — an incident traced to its mechanism. An untraced failure is not a lesson yet.
3. Wins — patterns that demonstrably worked and would not be rediscovered cheaply.

Mine the real record, not memory: the session ledger, the conversation (the user's messages above all), per-lane artifacts, PR review threads.

Where each source lives:

• The session ledger — in the scratchpad directory listed in the system prompt.
• The full conversation — a JSONL transcript under the Claude config directory at projects/<cwd with slashes replaced by dashes>/<session-id>.jsonl; the current session is the most recently modified file there, and compaction summaries cite the exact path. Long sessions grow to hundreds of megabytes — never read one wholesale. Extract layers with a script filtering by role: user messages and hook feedback first (highest signal, smallest volume), assistant narration only if needed, then slice for miners.
• Per-lane artifacts — subagent transcripts live beside the session transcript; lane ledgers live in the scratchpad.
• PR review threads — gh pr view <number> --comments and gh api repos/<owner>/<repo>/pulls/<number>/comments.