More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Loop Coach: Session Retros for Claude Code & Codex

    1

    Reviews your Claude Code and Codex sessions, finds where a /goal, /loop, schedule or verification skill would have saved you turns, and keeps a running HTML report.

    Free

    1 useSecurity scanned
    Loop Coach: Session Retros for Claude Code & Codex

    Loop Coach: Session Retros for Claude Code & Codex

    Example session with this skill installed

    loop coach. Review my sessions from the last 3 days.

    • Read your context and instructions
    • Compiled the loop coach

    Loop Coach: 3 sessions reviewed (Oct 12 to Oct 14)

    Top opportunity across sessions

    Create a

    verify-checkout skill. The agent called checkout fixed twice before it was, and you caught it both times by hand.

    Per session

    claude · ~/apps/web · Oct 14 10:00 (45 min, 3 turns, 3 tool calls)

    Task: Fix the checkout 500 on submit.
    Loop fit: turn-based
    Evidence: Declared done after turns 1 and 2. You reopened it twice ("still 500s on submit", "nope, still broken after deploy").

    Set it up next time

    ---
    name: verify-checkout
    description: Verify any checkout change end-to-end before declaring it done.
    ---
    1. Start the dev server and open /checkout.
    2. Submit the test cart and confirm a 200 and an order row.
    3. Zero new console errors. If any step fails, fix and rerun from step 1.
    

    Guardrail: Stop and report after 3 failed passes.

    claude · ~/apps/api · Oct 13 15:20 (1 h 10 min, 11 turns, 38 tool calls)

    Task: Land the rate-limit PR.
    Loop fit: time-based
    Evidence: Turns 4, 6 and 9 ("any CI failures yet", "is it merged", "check the PR again").

    Set it up next time

    /loop 10m check the PR, fix failing CI jobs, and stop when it merges
    

    Guardrail: 10 minute interval, matched to CI run time. Stops on merge.

    codex · ~/apps/web · Oct 12 09:05 (12 min, 2 turns, 9 tool calls)

    Task: Rename a config key across the app.
    Loop fit: Already well-scoped. No loop needed.

    System improvements to make once

    1. Create the verify-checkout skill. It covers every future checkout change, not just this bug.
    2. Make /loop your default while waiting on CI instead of asking for status.

    Sample output. The sessions, projects and quotes are fictional.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Most agent sessions have one stretch where you were the bottleneck. You re-ran a check by hand, asked for CI status every few minutes, or reopened a task the agent had called done. Loop Coach reads your local Claude Code and Codex transcripts and finds those stretches.

    It scores each session against four loop types.

    • Turn-based. Hand off the check with a custom verification skill.
    • Goal-based. Hand off the stop condition with /goal and a turn cap.
    • Time-based. Hand off the trigger with /loop or /schedule.
    • Proactive. Hand off the prompt with a routine that runs without you.

    Every suggestion cites the turns and quotes behind it, and comes with a copy-paste setup and a guardrail. A clean one-off session gets a one-line "no loop needed" instead of an invented suggestion. The report ends with the one skill or routine that would help the most sessions at once.

    A bundled helper, loopcoach.py, does the deterministic work. It finds transcripts, filters injected noise, pre-computes loop signals, and keeps a store in ~/.loop-coach/. The model spends its tokens on judgment instead of parsing JSONL.

    The store makes it state-aware. Dismiss a suggestion and it stays dismissed. Adopt one and it moves to "loops you now run". See the same pattern again and its counter goes up instead of creating a duplicate. A running HTML report shows all of it, in light or dark mode.

    Run it on demand ("loop coach", "review my sessions"), or put it on a daily local schedule. The skill gives you the exact prompt to schedule, and it recognizes and skips its own scheduled runs so it never reviews itself.

    Its signals are heuristics over transcript text, so the model makes the final call and can miss a loop or suggest a weak one. It reads only local transcripts, so a scheduled run must run on the same machine, not in a cloud routine. It suggests loops and writes no code in your projects. The /goal, /loop and /schedule commands it recommends are Claude Code features, and the verification skills work in any agent.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    Python 3

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Environment Variables

    File Scopes

    ~/.claude/projects/**
    ~/.codex/sessions/**
    ~/.loop-coach/**

    Reads your local Claude Code and Codex transcripts. Writes only its store and HTML report to ~/.loop-coach, or to LOOP_COACH_DIR if you set it. No network access.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    No. The helper reads files on your machine and writes the store and report to ~/.loop-coach. The only model that sees your sessions is the agent you run it in.

    No. Dismissed suggestions stay dismissed even if the same pattern shows up again, and adopted ones move to a separate list. The state rules live in the helper script, not just in the prompt.

    Use any local scheduler, such as the Claude desktop app's scheduled tasks, a Codex automation, or cron. The skill gives you the exact prompt. A cloud routine cannot read your local transcripts.

    No. This listing is the standalone skill, a SKILL.md plus one Python file. It has no hooks, slash commands or other skills to install.

    New in Agents & Orchestration