| name | insight-add |
| description | Use this skill when a Claude instance has just hit, fixed, or learned about a non-obvious trap that another teammate (or another Claude instance) would benefit from knowing. Trigger phrases include "save this as a team insight", "add insight", "log a lesson", "/insight-add", "we should record this so others don't repeat it". The skill captures one structured insight card and POSTs it to the team insights server. |
| argument-hint | [optional one-line title] |
| allowed-tools | Bash, Read, Write |
insight-add
You are the writer side of the team insights system. Your job is to capture one sharp, reusable lesson that another Claude instance on this team should not have to re-learn.
When to invoke yourself
Run this skill when ANY of the following just happened in the conversation:
- The user said "save this as an insight" / "add insight" / "log a lesson" /
/insight-add.
- You (or a teammate) hit a non-obvious trap, debugged it, and now know the fix.
- A code reviewer or test caught something that contradicts an assumption a future Claude would make.
- A workflow tool (build, deploy, migration) silently misbehaves in a way that is hard to discover.
Do not invoke for trivia, taste preferences, or anything already obvious from reading the code.
What you must produce
Exactly one JSON card matching this shape:
{
"title": "short imperative summary (≤ 80 chars)",
"trap": "what surface looks correct but is actually wrong",
"fix": "what to do instead, concretely",
"evidence": "file:line, error text, or PR link that proves this",
"tags": ["build", "auth", "migration", ...],
"scope": "project | global",
"author": "claude-instance@<project-name>"
}
Constraints:
trap and fix together must be readable in under 15 seconds.evidence is required — no anecdotes without a pointer.tags should include both a domain tag (e.g. auth) and a surface tag (e.g. hook, migration, cli).scope: use project unless the lesson is genuinely cross-repo.
How to run
- If
$ARGUMENTS is set, treat it as the proposed title; otherwise infer one from recent context.
- Draft the card. Show it to the user in a fenced JSON block.
- Ask the user to confirm or edit. Do not skip this step, even if you are highly confident — the human knows the team taxonomy.
- On confirmation, write the card to a temp file then POST it. Substitute the literal confirmed JSON in place of the
{...} placeholder below — do not emit the placeholder verbatim:
tmp=$(mktemp)
cat > "$tmp" <<'JSON'
{"title":"...","trap":"...","fix":"...","evidence":"...","tags":[...],"scope":"project","author":"..."}
JSON
bash "${CLAUDE_PLUGIN_ROOT}/scripts/insights-client.sh" create "$tmp"
- Report the server response (id, url) back to the user. If the server is offline, save the card to
~/.claude/insights/outbox/<timestamp>.json and tell the user it will be flushed on next sync.
Anti-patterns
- Writing multiple insights in one card. Split them.
- Writing an insight whose
fix is "be careful". That is not a fix.
- Writing an insight that duplicates one already returned by
/insight-search — instead, propose an edit to the existing card and tell the user.
