Skills

confab setup ships two bundled skills — slash commands that get installed into your provider CLIs (Claude Code, Codex) and that you invoke from inside an active session. Both are pre-wired to talk to the backend you authed against during setup, so there’s no extra config.

How skills are installed

When you run confab setup, the CLI:

1. Auto-detects which provider CLIs (claude, codex) are on your PATH.
2. For each detected provider, writes the skill’s SKILL.md into the provider’s skills directory:
   • Claude Code: ~/.claude/skills/<skill>/SKILL.md
   • Codex: ~/.codex/skills/<skill>/SKILL.md
3. If you’d previously customized the skill, the existing file is backed up to SKILL.md.bak before being overwritten — so re-running setup never silently drops your edits.

You can install or remove skills manually with confab skills add / confab skills remove (see Commands).

The bundled skills are:

/til — Today I learned

Captures a short knowledge snippet from the current session and saves it to your Confabulous backend, where it surfaces in the TILs feed.

Usage

Inside an active Claude Code or Codex session:

/til <what you learned>

The agent:

1. Uses your message as the TIL title.
2. Writes a 2–3 sentence summary that captures what was learned and why it matters, drawing on the conversation context.
3. Invokes confab til under the hood to persist the TIL to the backend.
4. Briefly confirms back to you that the TIL was saved.

The TIL deep-links back to the exact message in the originating session, so future-you can click through to the full context.

Under the hood

The skill is allowed to invoke confab til *. Behind the scenes:

# Claude Code:
confab til --session "${CLAUDE_SESSION_ID}" --title "<title>" --summary "<summary>"

# Codex:
confab til --provider codex --session "${CODEX_THREAD_ID}" --title "<title>" --summary "<summary>"

For Codex, the CLI normalizes subagent thread IDs to the root thread before saving — so TILs from a subagent attach to the parent session.

/retro — Session retrospective

Fetches and discusses a session you (or a teammate) ran earlier. Useful for reviewing what happened, extracting learnings, or critiquing the approach.

Usage

Inside a fresh Claude Code or Codex session:

/retro <session-id> [optional question or focus]

<session-id> is the ID shown on your Confabulous dashboard. The optional trailing text steers the discussion.

The agent:

1. Calls confab retro to fetch the condensed transcript plus structured metadata.
2. Writes the JSON and XML output to a timestamped directory under /tmp/retro-<timestamp>/.
3. Searches for the local raw transcript file (richer than the backend’s condensed view) and reads relevant sections for deeper analysis.
4. Summarizes the session conversationally — what happened, key outcomes, cost, duration, model.
5. If you supplied a question or focus area, answers it; otherwise opens up to general discussion.

Under the hood

RETRO_DIR="/tmp/retro-$(date +%s)"
confab retro --output-dir "$RETRO_DIR" <session-id>

This writes two files (response.json and transcript.xml) into the output directory. The skill’s prompt is in SKILL.md and is what teaches the agent how to use them.

Customizing or replacing skills

The SKILL.md files are plain markdown — you can edit them in place. Your edits are preserved across confab setup runs (a .bak is written before overwrite).

To pin a customized version: edit SKILL.md, then run confab status to confirm the skill is still recognized as installed. To revert: delete the file and run confab skills add.

To author entirely new skills for your team’s workflow, drop a SKILL.md into the provider’s skills directory directly — Confabulous doesn’t restrict what skills can live there beyond its own bundled set.