OPEN-SOURCE REFLECTION SKILL

Learn from how you work with AI.

Models change, and the way you work drifts out of step. second look reads your sessions across Claude Code, Codex, Cursor, ChatGPT and more, finds the leaks in your prompts, context and skills, and gives you a few fixes to try. Each one cites the sessions it came from.

See a sample reportGet the skill

Last 12 months · 1,330 sessions

Corrections per 10 messages

How often you redirect or fix the AI. Lower means it understood you sooner.

1.2

Down from about 1.9

Briefed starts

52%About the same

Output checking

34%−11 pts

Restarts

15%+9 pts

In the last 4 weeks, output checks fell from 45% to 34% and restarts rose from 6% to 15%.

The opportunities

Where a closer look pays off.

  1. 01

    Less overhead for tasks

    Cut the setup and long-session overhead around each task, so work moves faster and your usage goes further.

  2. 02

    Better output, fewer errors

    Find the habits behind rework, so first results need less fixing.

  3. 03

    More reliable code

    Ask for code with a clear scope and a way to check it.

  4. 04

    Keep your own judgement

    Depend less on AI for the thinking.

  5. 05

    Deeper research

    Get more thorough research and sharper reasoning from each session.

  6. 06

    Skills that transfer

    Build skills that carry over to new tools and models.

  7. 07

    Bring tools together

    Get more from several AI tools by matching each task to the right one.

  8. 08

    Less repeated setup

    Stop retyping context. Move it into memory and skills.

Our approach

AI use is more than the prompt you type.

It also reads the context you give, your skills and config, and how your habits change. Each finding shows its evidence and one small change to try.

Sources it can’t open are marked “not checked.”

WHAT THE SKILL LOOKS FOR

Did your last fixes work, and did your first message give the goal, audience, constraints and what done looks like?

WHAT YOU GET BACK

WHAT HAPPENS
Your prototype requests name the feature but leave out the user.
EVIDENCE
In three sessions, that cost several turns of clarifying.
TRY THIS
Add who it’s for and what done means to the first message.

Generic advice like “write clearer prompts” doesn’t make the cut.

Which skills or configs are stale, broken, duplicated, or loading into every session unused?

Are your habits improving, and where did your usage go?

Every two weeks

Review, plan, check back.

After your first report, second look offers to run the review every two weeks for you, or at the cadence you pick. Change it or turn it off any time by asking your assistant.

  1. 01

    Coach this cycle

    Get a fix to test for every pattern.

  2. 02

    Maintain your AI setup

    Flag stale, broken or duplicate skills and config.

  3. 03

    Review change over time

    Compare habits and outcomes across tools.

  4. 04

    Return to the evidence

    Did it stick, did it help, and what did it cost?

Privacy

What it reads, and where it goes.

second look has no server and no account. Nothing about your sessions is sent to us.

  1. 01

    Your files

    ~/.claude, ~/.codex, Cursor’s database, ~/.gemini, and the exports you add. Read only when you ask for a review.

  2. 02

    Your assistant

    Claude Code, Codex or Cursor runs the review on your computer and writes the report to ~/.second-look.

  3. 03

    Your model provider

    Sees what the review reads, like any message you send your assistant, under your existing account settings.

  • What it reads

    Claude Code, Codex, Cursor and Antigravity logs on this computer, the exports you add, and your skill files. Run install.sh --check to see the list for your machine.

  • What gets removed

    API keys, emails, and token or password values are stripped before anything is quoted. It works by pattern, so read a report before you share it.

  • Who sees the report

    Only you. It’s saved in ~/.second-look or published as a private artifact, and it’s shared only if you choose to.

  • Delete it any time

    Stop recurring reviews, then remove ~/.second-look and the skill folder. If you kept a year of Claude Code history, set it back to 30 days in one setting.

Using AI tools at work? Check your company’s AI policy before you review work sessions. Keeping a year of history also means more of it stays on your disk; the installer asks before changing that.

Privacy details on GitHub →

Coverage

Supported AI tools

Start a review from Claude Code, Codex or Cursor, and one run covers every source below that’s on this computer.

ToolWhat’s readSetup neededHow far back
Claude CodeSession logs in ~/.claude/projects, your skills and CLAUDE.mdNone. The installer links the skill inLast 30 days only. Claude Code deletes older sessions unless you raise cleanupPeriodDays in ~/.claude/settings.json
CodexSession logs in ~/.codex/sessions, your skills and AGENTS.mdNone. The installer links the skill inFull history. Nothing to change
CursorChats in Cursor’s local database, plus .cursor/rulesNone. The installer links the skill inFull history. Nothing to change
AntigravityAgent transcripts in ~/.gemini/antigravity-cli, plus ~/.gemini/GEMINI.mdNone. Reviews start from one of the tools aboveWhatever the CLI keeps on disk. Support is new, and app chats may be missing
claude.aiconversations.json from your data exportExport it from Settings → Privacy, then ask your assistant to “add my claude.ai export”Optional. Covers everything up to your export date
ChatGPTconversations.json from your data exportExport it from Settings → Data controls, then ask your assistant to “add my ChatGPT export”Optional. Covers everything up to your export date
GeminiMyActivity.json from a Google Takeout export of your Gemini Apps activityIn Google Takeout, pick My Activity → Gemini Apps, format JSON, then ask your assistant to “add my Gemini export”Optional. Covers chats saved with Gemini Apps Activity on, up to your export date

Sample report

Oct 25, 2026

Sample data

Over the past year, you’ve gotten better at explaining what you want, and the AI misreads you less. Lately the Streak redesign pulled the other way: you’ve been checking prototypes less and starting over more often.

Improve your prompts and setup

Vague prototype starts take the most time. Syncing your rules is the quickest fix, so it comes first. Five fixes below, all in your plan until Nov 8.

Did the fixes from your last review work?

  • Keep

    Confirm the plan before editing. You’re correcting the AI about as often as before, at 1.2 per 10 messages. The fix is only 2 weeks old, too soon to tell.

  • Keep

    Send research comparisons to a helper agent. Sessions using one rose from 11% to 18%, but sessions crowded with lookups stayed about the same.

  • Revise

    Brief before big prototype tasks. You haven’t tried this yet, so it’s smaller now: only longer prototype sessions. It’s fix 2 in your plan.

  1. 1

    You retype the same rules across your tools

    Try this: Add your rules to the tools missing them

  2. 2

    Prototype sessions start vague, then restart

    WHAT HAPPENS

    Your first message names what to build, but not who it’s for or what done looks like. You fill in the gaps over the next few messages, and sometimes start a new session.

    EVIDENCE

    In the last 4 weeks, 16 of your 25 restarts were prototype sessions, including 4 Streak settings-screen sessions on Oct 14–15. In those 4 weeks, 44% of prototype sessions opened with a brief.

    WHY IT MATTERS

    Each restart repeats the setup, and without a definition of done it’s hard to tell when a prototype is finished.

    Try this: Brief before big prototype tasks

    Saves about 30 minutes a week

    Starting a prototype that’s more than a quick tweak? Open with the goal, the user, what done means and the limits. If you don’t know what done looks like yet, say so.

  3. 3

    Prototypes often go out without a last check

    Try this: Check beta work before it ships

  4. 4

    Long sessions drift across projects and eat your usage

    Try this: Start fresh between projects

  5. 5

    Your strongest model does your simplest work

    Try this: Match the model to the job

Your plan until Nov 8

FIXTIMESAVES
Add your rules to the tools missing them5 minutes, onceAbout 5 minutes a week
Brief before big prototype tasksA minute before any prototype that’s more than a quick tweakAbout 30 minutes a week
Check beta work before it ships5 minutes per piece others will seeFewer fixes after launch
Start fresh between projectsA habit, whenever you switch projectsAbout a tenth of your Claude Code and Codex usage
Match the model to the jobA few seconds per sessionAbout a tenth of your Claude Code usage

Keep your skills current

Three of your seven skills give wrong information today.

Checked Oct 25: 7 skills you wrote, 17 you installed. Couldn’t read your Cursor skills folder, so any skills there aren’t counted.

Skills to fix

ISSUEFIX
pr-review names a tool that doesn’t exist, so the visual check can get skippedPaste the fix (5 min)
release-notes still lists last quarter’s pricing tiersPaste the fix (5 min)
beta-feedback-synth says the team has 4 engineers; your notes say 6In claude.ai, paste: “Change the team size in my beta-feedback-synth skill from 4 to 6.” (1 min)

How your habits are changing

A full year of claude.ai; Claude Code since Jun 29; Codex since Apr 20; Cursor since Feb 9; a full year of ChatGPT. Gemini isn’t added.

Last 4 weeks compared with the 4 before: you restarted more tasks and checked the work less often.

Briefed starts

52%About the same

NovOct

Share of sessions where your first message gives the goal, the context and what done looks like.

Corrections / 10 msgs

1.2About the same

NovOct

How often you redirect or fix the AI, per 10 messages. Lower means it understood you sooner.

Output checking

34%−11 pts

NovOct

Share of sessions where you test, review or question the output before accepting it.

Restarts

15%+9 pts

NovOct

Sessions that start over on a task you began in the last 2 days.

Getting started

Try it on your own conversations.

One install covers Claude Code, Codex and Cursor. Everything is read on your computer, and nothing is uploaded.

Get second look on GitHub
  1. 1

    Ask your assistant to install it

    Paste this into Claude Code, Codex, or Cursor. It tells you which folders it will read before it changes anything.

    ›Install second look from github.com/txz8096/second-look.

    Prefer the terminal?

    curl -fsSL https://raw.githubusercontent.com/txz8096/second-look/main/install.sh | bash
  2. 2

    Add your claude.ai, ChatGPT and Gemini history

    Optional, and you can add it any time. In claude.ai, go to Settings → Privacy → Export data. In ChatGPT, go to Settings → Data controls → Export data. For Gemini, use Google Takeout: My Activity → Gemini Apps, format JSON. Each sends you a zip, usually within a day. Download it, then ask your assistant. It finds the zip in Downloads, or you can drag it into the chat:

    ›add my ChatGPT export
  3. 3

    Ask for a second look

    Ask your assistant for a second look, or type /second-look in Claude Code. The first run asks four questions, then builds a private report you can check and correct.

    ›/second-look

Claude Code only? Install the plugin instead: /plugin marketplace add txz8096/second-look, then /plugin install second-look@second-look. On claude.ai? Download second-look.zip from the latest GitHub release, turn on code execution in Settings → Capabilities, and upload it under Skills.

FAQ

Do I need to install it in every tool?

No. The installer puts one copy on your computer and links it into Claude Code, Codex and Cursor, whichever you have. A review runs in the tool where you start it, and it reads every tool’s files from there.

Which AI apps does it support?

Seven. Claude Code, Codex and Cursor can run a review and are read directly. Antigravity is read directly but can’t start a review, so start one from another tool on the same computer. claude.ai, ChatGPT and Gemini (the web app) are read from data exports you download, so they count up to the day you export. Gemini’s export has no conversation boundaries, so its sessions are grouped by time and its numbers are estimates. Antigravity support is new and built on an undocumented file format, so a release that changes it can hide your sessions until second look is updated.

Does one report cover all my AI tools?

Yes, if you start it from Claude Code (terminal or desktop app), Codex or Cursor on the computer where you use them. One run reads their logs and Antigravity’s, plus claude.ai, ChatGPT and Gemini if you add their data exports. Started from claude.ai in a browser or from a cloud session, it can’t see those logs, so it covers only your claude.ai chats.

What data does it read?

Only what you point it at: local Claude Code, Codex, Cursor and Antigravity logs, the claude.ai, ChatGPT and Gemini data exports you download yourself, and your skill and config files for the upkeep check. Cloud sessions aren’t included.

What do I need to run it on claude.ai?

Download second-look.zip from the latest GitHub release, turn on code execution, chat search and memory in Settings, and upload the zip under Skills. Without chat search or a conversations.json export, habit numbers show as “Not measured” instead of a guess.

What happens the first time I run it?

It asks four short questions: what you build, which areas to watch for new releases, where your standing rules live, and where your skills are. Your answers become a profile in ~/.second-look/profile.md that you can edit later.

How often should I run it?

Every two weeks works well. Each review checks last cycle’s experiments, then gives a fix for every new pattern. You can also ask for a review any time; it covers everything since the last one. Name a narrower scope, like “my Cursor sessions”, if that’s all you want.

Does it run by itself? Can I change or stop it?

Yes, if you say yes after your first report. In the Claude desktop app or the Codex app, a scheduled task runs the full review on your computer. In a terminal, a small daily job runs it in the background and tells you when the report is ready. In Cursor or Antigravity, which can’t schedule anything on your computer, you get a notification when a review is due instead. Every scheduled review includes Antigravity sessions and your exports, and the notification tells you when an export is over 90 days old. Each run uses your quota and needs your computer on. Ask your assistant to change the cadence (“make it monthly”) or stop it (“stop the reviews”).

Which apps can run it on a schedule?

It depends on whether the app can run the review on your computer, because that’s where your session logs are.

Where you use itRuns by itself?How it recursNeeds
Claude desktop appYesA scheduled task in the app runs the full reviewThe app open, computer on
Codex appYesA scheduled task in the app runs the full reviewThe app open, computer on
Terminal (Claude Code or Codex CLI)YesA daily job runs the review in the background, then tells you the report is readyComputer on; file access approved in advance
CursorNoA notification when a review is due; you start itComputer on (macOS or Linux)
AntigravityNoA notification when a review is due; start it from Claude Code, Codex or Cursor. Any scheduled review includes your Antigravity sessionsComputer on, and one of those tools installed
claude.aiYes, claude.ai onlyA claude.ai scheduled task. It can’t see your coding toolsChat search on
ChatGPTNoReviews don’t run here. Its history comes in through the data exportA conversations.json export
Gemini (web app)NoReviews don’t run here. Its history comes in through the Takeout export, and you’re reminded when it’s 90 days oldA MyActivity.json export

How far back does the yearly view go?

As far as your logs do. Claude Code keeps 30 days by default, which is why the installer offers to keep a year. claude.ai, ChatGPT and Gemini count up to the day you export. Each run saves weekly totals, so weeks your tools later delete stay in the trend. The sample report on this page uses made-up data.

Does it show where my token usage goes?

For Claude Code and Codex, the only tools that log tokens. The charts track how much of your usage came from very long sessions, where every request re-reads a long history. The skills check names what loads into every session before you type, like skills you haven’t run in 60 days, and how much you’d save by having them run only when you ask. Numbers compare you with your own typical session, never with other people. A full breakdown appears only if you keep hitting usage limits.

How accurate are the findings?

They’re estimates read from the wording of your sessions. Each finding shows its examples and how sure the review is (Sure, Likely or Hunch). Before trusting a number, the review spot-checks a sample of its own labels, and you can correct a label at the next review.

Where does it run, and is anything uploaded?

On your machine, and the report is saved in ~/.second-look. Nothing goes to second look, because there’s no second look server. The review runs inside your assistant, so what it reads goes to your model provider like any other message you send, under your existing account settings.

What is redacted?

Prompts are trimmed, and API keys, emails, and token or password values are removed before anything is quoted. This works by pattern, so it can miss things like a client’s name. Read a report before you share it.

Can I edit or delete the report?

Yes. The report, your profile, the review log and your history are plain files in ~/.second-look on your computer. Edit or delete them whenever you like.

How do I uninstall it?

Ask your assistant to stop the reviews, if you set them up. Then delete ~/.agents/skills/second-look and the second-look links in ~/.claude/skills, ~/.codex/skills and ~/.cursor/skills. Delete ~/.second-look too if you don’t want to keep your reports. If you used the plugin, run /plugin uninstall second-look.