tunelisted
Install: claude install-skill duthaho/skillhub
# tune — the harness retro
`/tune` — audit the current project's harness against recent sessions
`/tune all` — widen across every project under `~/.claude/projects/`
Answer one question: **what keeps going wrong across sessions, and what
one-line change would make it impossible next time?** Correcting the agent
fixes one session; correcting the harness fixes every session after it.
The evidence is already on disk — the transcripts recorded every
correction, every retyped instruction, every permission prompt. This skill
reads them so the user doesn't repeat them.
## Step 0 — Scope, then inventory
**Scope:** default is the current project's transcripts — the last ~10
sessions or 2 weeks, whichever is smaller — from
`~/.claude/projects/<project-dir>/*.jsonl` (the dir name is the project
path with `/` → `-`). `/tune all` widens to every project; mining gets
shallower per project, say so.
**Inventory the harness first** — findings only mean something against the
baseline. Read: user + project `CLAUDE.md`, the installed skills (names +
descriptions), `settings.json` / `settings.local.json` (permissions,
hooks). Also read `out/tune/learnings.md` (past accepted/rejected — don't
re-propose what was rejected), `out/tune/principles.md` if present
(promoted cross-cutting principles — a lens for judging new findings),
and `out/dev/bugfix-log.md` if present (recurring root causes are harness
findings too).
## Step 1 — Mine the transcripts (cheap models, parallel)
Bulk transcript rea