Agent Wrapped

Run everything through wrapped.sh next to this file. On first use it downloads the latest agent-wrapped release for this OS and CPU from GitHub into ~/.local/share/agent-wrapped (a 40 to 60 MB download), and it checks for a newer release at most once a day, so never install anything by hand. The first video on a machine also downloads Chrome Headless Shell, so allow a few minutes for it.

On Windows, wrapped.sh works from Git Bash. From PowerShell with no sh, run powershell -ExecutionPolicy Bypass -File <skill-dir>/wrapped.ps1 with the same arguments everywhere below; both share one install.

Everything is read and rendered locally. Nothing is uploaded.

1. Read the history

<skill-dir>/wrapped.sh stats -o "${TMPDIR:-/tmp}/wrapped.json"

It finds every supported agent with local history, ranks them by session count, and prints a one-line summary. Add --year 2025 for another year and --name "..." if the user gives a name for the first card. The JSON holds prompt excerpts, so keep it in the temp directory, never in a project.

T3 Code is an app that runs these harnesses, not an agent: each T3 thread counts once, as the harness that ran it.

Other machines

When the user wants every machine counted, export each one's sessions over SSH (it installs the binary there the same way) and add them:

ssh <host> 'curl -fsSL https://raw.githubusercontent.com/vkpdeveloper/agent-wrapped/main/skill/agent-wrapped/wrapped.sh | sh -s -- sessions --year <year>' > "${TMPDIR:-/tmp}/wrapped-<host>.json"
<skill-dir>/wrapped.sh stats --add "${TMPDIR:-/tmp}/wrapped-<host>.json" --add ... -o "${TMPDIR:-/tmp}/wrapped.json"

Use only hosts the user named or already reaches by SSH alias. Exports mask secrets and show the home directory as ~. A session found on two machines counts once.

2. Pick the video

Unless the user already said which one, offer the menu: for each format, what it looks like, with their own numbers from the JSON as the hook. Recommend the one whose numbers are most extreme, then stop and wait for them to pick one or several. Never pick for them unless they told you to.

When they named a format, or told you not to ask, make that one, and in the report list the other formats their numbers suit, a line each with its hook, so they can ask for more. Nothing is recounted: each extra format is its own jokes (step 5) and one render.

previews/<cut>.jpg next to this file is a frame of each format. If you can show images, show the previews of the formats you offer.

--cut	What it looks like	Music	Hook from the JSON
full	The whole year, about 50s, Spotify Wrapped style: big numbers on green, agent bars with their logos, projects, a 3am clock, lessons, persona	Future bass	totals.hours, agents[0].name
social	30s, first person ("I spent 724 hours"): giant numbers on black and green, their own lines, ends on "What's yours?"	Big room EDM	totals.hours, persona.title
rage	A desk on fire, a swear counter, a censored leaderboard of their words, the peak month, air horns	Rage trap	swears.total, swears.words
loop	"I told my agent no 16 times": their pushback from that session as chat bubbles, then the retries	Glitch hop	loop.corrections, loop.retries
credit	Commits and pull requests, then bars of which agents co-signed them	Heist jazz	git.commits, git.coAuthors
grind	A city at night: the longest day, the streak, the latest night	Dark synthwave	busiestDay.hours, streak.days
boss	A penthouse office: the agents they manage, the tool calls, a graded performance review per agent	Mafia trap	agents.length, totals.toolCalls
speedrun	Pixel art and a speedrun timer: splits, skips used, deaths	Chiptune	totals.hours, totals.failedCommands
genz	A loud brainrot backdrop, huge captions one word at a time ("lowkey my roman empire", "aura")	Phonk	swears.please, corrections.total
linkedin	A fake LinkedIn post: "I'm humbled to announce", a stock photo of a man shaking a robot's hand, "Agree?"	Corporate ukulele	totals.sessions, corrections.total
docu	A jungle, a giant faded number, nature documentary subtitles about them	Orchestral	lateNight, swears.words
news	A TV studio: LIVE bug, a big number in a box, breaking news chyron, scrolling ticker	News theme	swears.total, loop.retries
dating	A sunset, their agents as dating profile cards: the main one, the ex, the talking stage, red flags	R&B slow jam	agents (needs 2 or more)
tier	An S to D tier list, agent logos dropping into the rows	Hard trap	agents (needs 3 or more)
receipt	A printed receipt on a counter, a line per stat, the persona as the total, a barcode	Elevator bossa	totals
trailer	A stormy city, spaced serif title cards ("In a world where agents write the code"), the persona as the title	Trailer braams	totals.hours, persona.title
chat	Their agents' group chat gossiping about them, logos as avatars	Lofi	agents (needs 2 or more), swears.please
convo	Their real messages with their agents, one per screen like a chat screenshot: their line types and sends, the agent's reply lands. Any language	Comedy hip hop	exchangeSamples (pick convo), totals.prompts

Leave out a format whose numbers are zero or tiny (2 swears is not dev rage). Every format but full is made to post: no projects, tools or models, and swears are censored on screen. All are about 13 to 30 seconds.

3. Write the lessons

Read corrections.samples in the JSON: real prompts where the user pushed back on their agent. Group them by what the agent kept getting wrong and write up to 3 lessons, most frequent first:

• rule: an instruction the user could paste into CLAUDE.md or AGENTS.md, at most 8 words, imperative and concrete ("Use bun, never npm", "Run the tests before saying done", "Ask before deleting files").
• count: how many of the samples that rule covers. Count honestly; the video prints it.

Skip one-off complaints and anything that names a person, client or secret. Leave lessons empty if no pattern repeats. You may sharpen persona.line, but keep its number exactly as computed.

4. Pick the quotes

Read promptSamples: up to 200 short prompts the user typed, spread over the year. Get a feel for how they talk to agents (blunt, polite, all caps, typos, 3am despair, tiny wins), then copy 3 or 4 that sound most like them into quotes.lines, keeping each one's date, time and agent:

• Pick lines that are funny, oddly specific or very them. A stranger should get the joke without context. Mix moods; skip boring task descriptions.
• Keep their words, typos and casing. You may cut a quote down to its best part, at most 120 characters, but never add words.
• Skip anything with a secret, a ********, a person's or client's name, or anything they would not want posted. With --blur-projects, also skip quotes that name a project.

Write quotes.title: at most 6 words, in the user's own tone, introducing the quotes ("Things you said at 3am", "Your agent has receipts"). Leave quotes.lines empty if nothing is worth posting; the card is then left out.

Only for the formats they picked:

• loop: loop.lines holds that session's pushback. Reorder it so the two funniest come first (the video shows two), by the same rules as the quotes.
• convo: read exchangeSamples, things they typed (text) with the agent's first words back (reply), and copy the 3 or 4 funniest pairs into convo, funniest first, keeping date, time and agent. The reply is the joke: an agent owning up ("Yes, I broke the status format."), "You're absolutely right", a deadpan answer to a rant. Cut the reply down to its funny part, at most 100 characters, and the text to at most 120, but never add or translate words: Hindi, Hinglish or any other language shows as typed. Same skips as the quotes. Leave convo empty to skip the cut.
• boss: write reviews, a performance review for each of the top 4 agents in agents: name, a grade ("A+", "B-", "C") and a line of at most 6 words, a joke grounded in its real numbers ("969 sessions. Never sleeps."). Leave it empty to skip the card.

Save the edited JSON. Lessons only show in full, quotes in full and social, and convo in its own cut, so skip the steps a picked format does not use.

5. Write the jokes

Every format but full is a meme, and the jokes on it should be about this user, not anyone's. The user should not have to explain the joke: the brief below carries it. For each format they picked, print its brief:

<skill-dir>/wrapped.sh lines "${TMPDIR:-/tmp}/wrapped.json" --cut <name>

Write the jokes from the brief alone; there is no need to read the JSON again.

• Angle: the format's comedy. Every line plays it.
• Facts: their numbers, already worked out: who they cheat on, who they ghosted, their worst hour, the pleases against the swears.
• Their words: their loudest prompts and the times an agent owned up, verbatim. A line they really typed beats any line you can write, so quote them (you may cut one down, never reword it).
• Cards: each card that takes a joke, what it already shows, and its limit.

Write one into lines, keyed by card id ("lines": { "rageHook": "..." }). A card left out keeps its built-in joke, so skip any you cannot beat. For chat, the thread goes in chat as { "from": "<agent id>" | "me", "text": "..." }.

• Write like a meme page, not a report: roast them and their agents, be specific and a little unhinged, end on the punch. Use their own slang and language (Hinglish stays Hinglish).
• Never invent or change a number; the video prints what you write.
• Call the user "I", "me" or "they", never "he" or "she": the history does not say who they are.
• No project, client or person names, no secrets, nothing they would not post. Swears are censored on screen anyway.

Then check every line and rewrite each one that fails:

1. It holds a number from Facts or words from Their words.
2. It would be wrong about any other user. A line that fits everyone is a miss.
3. It adds what the card does not already say: a sticker under "350 times" brings a new fact or the punch, not the 350 again.
4. A stranger would laugh, or wince, without context.
5. No two lines lean on the same quote or fact; each card gets its own joke.
6. It is within its limit.

6. Render

<skill-dir>/wrapped.sh render "${TMPDIR:-/tmp}/wrapped.json" -o ~/Desktop/agent-wrapped-<year>.mp4

Use ~/Desktop when it exists, otherwise the current directory, unless the user names a place. --format square (1080x1080) for the X or LinkedIn feed; the default is vertical (1080x1920) for Stories, Reels and TikTok. Add --blur-projects when the user wants project names and session titles hidden. Add --cut <name> for the format they picked, and render each one they picked to its own file (agent-wrapped-<year>-<cut>.mp4); the JSON is read again each time, nothing is recounted. Every format has its own music track, on the same beat as the cuts, with sound effects on every hit.

Report

Give the video paths, the other formats worth making (see step 2), the headline numbers (hours, sessions, the #1 agent), the persona, the lessons and the quotes you picked. Offer to add the lessons to their CLAUDE.md or AGENTS.md. Remind them that full shows project names (unless blurred), and every format can show lines they typed, so they should watch it before posting.