More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Retro: Turn Session Mistakes Into Harness Fixes

    1

    Fixing a mistake once does not stop Claude from repeating it next week. Claude reads its own transcripts for loops, false "done" claims, and repeat corrections, and turns each into a hook, skill edit, or rule.

    Free

    Security scanned
    Retro: Turn Session Mistakes Into Harness Fixes

    Retro: Turn Session Mistakes Into Harness Fixes

    Example session with this skill installed

    Do a retro on this session. I had to correct you three times about the test command.

    • Read your context and instructions
    • Compiled the retro

    Claude scans the transcript and finds three signals: npm test run four times with the same failure, two user corrections saying the repo uses pnpm -w test, and a claim that tests passed that the next message contradicted. It adds a CLAUDE.md line naming the right test command and a hook that rejects npm test in that repo, so neither needs to be remembered. It writes no memory note because both are enforced, and reports each change with the file it touched and the failure it prevents.

    passed that the next message contradicted. It adds a CLAUDE.md

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

    About this skill

    About this skill

    A Claude Code session ends with a fix, and the next session makes the same mistake, because nothing in the setup changed. The lesson the skill is built on: fix the harness so the failure cannot recur, not the single mistake.

    The skill reads the current session or a transcript you name and looks for five signals: the same tool call repeated with no change, a claim of finished work that was later contradicted, a user correction, an approach abandoned after several tries, and an environment fact learned the hard way. Each signal becomes one rule. The skill first tries a structural fix (a hook, a skill edit, a CLAUDE.md line, an allowlist entry, or a wrapper script) so the failure becomes impossible, and writes a memory note only when no structural fix exists. Sweep mode, /retro sweep, clusters failures across every transcript since the last sweep, ranks a single failure that reached a person or cost money above frequent minor ones, and flags sessions over $50 or over 30% of the window's spend using ccusage.

    The proof

    There is no eval score. The SKILL.md records the measured traps it guards against, each from a real sweep: 17% of tool results in an mtime-selected corpus predated the window, so it filters on each event's own timestamp; 24 of 53 correction hits in one week came from skill-eval runs, so it drops repeated first prompts; 55 chained sleep calls looked blocked when only 2 of 55 were denied, so it matches each result to its call before counting a failure; and a ccusage --until set to today hid 3 sessions and 11% of a window's spend, so it asks for tomorrow.

    What you get

    One SKILL.md with the per-session procedure, sweep mode, the cost check, and the memory-writing rules. The headless weekly runner (bin/retro-sweep.sh) and a macOS schedule template are in the GitHub repo, not this package.

    How to use

    1. Copy the skill folder to ~/.claude/skills/retro/.
    2. At the end of a hard session, say "do a retro" or run /retro.
    3. Review the hook, skill, or CLAUDE.md edits it made in git, and run /retro sweep weekly, or schedule the repo's runner.

    Source and license

    Free to use and MIT licensed. The full source, install steps, and issue tracker are on GitHub at https://github.com/robertmnyborg/skill-eval.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Built for Claude Code's transcript and memory layout. The weekly runner, schedule template, and optional signal tables from the claude-oak-agents repo are not in this package. Some commands use macOS date syntax. Headless runs may be denied writes under ~/.claude, and the skill describes a workaround. It edits your harness files, so review its changes.

    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 buying a skill.

    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

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 8 days ago

    • Free with an Agensi account

    Software
    Claude CodejqPython 3Node.jsccusage

    Permissions

    Allowed Hosts

    registry.npmjs.org (npx ccusage)

    File Scopes

    ~/.claude/projects/
    ~/.claude/history.jsonl
    ~/.claude/settings.json
    ~/.claude/skills/
    CLAUDE.md
    files
    ~/.claude/.retro-sweep-last

    Reads your local session transcripts and edits harness files. The ccusage cost check runs through npx.

    Listed8 days ago

    What's inside

    Frequently Asked Questions

    Claude Code can save memory notes, but it does not go back through its own transcripts looking for loops and corrections, and a memory note is something the model must remember. This skill prefers changes the harness enforces, such as a hook, and writes memory only as a fallback.

    No. It reads Claude Code's transcript files under ~/.claude/projects and writes Claude Code memory, hooks, and CLAUDE.md rules.

    It edits hooks, skills, CLAUDE.md, and memory files as part of the fix. Those are files you can review and revert in git, and the report lists every change.

    No. `/retro sweep` runs on demand. The runner in the repo only makes it run headless on a schedule.

    On GitHub at https://github.com/robertmnyborg/skill-eval, under the MIT license. Every file in this package is there, plus the README and install steps.

    New in LLM Ops

    Free