name: daily-team description: Shows today's creative team and task, drawn from the date so they change every day. Has the team argue and decide on a plan for a brief, or builds the plan as single-file HTML artifacts. Use when the user asks for today's team or task, a creative team, a brainstorm with distinct roles, or new angles on a brief. argument-hint: "[publish | recruit | build [1-4]] [brief]" allowed-tools: Bash(sh team.sh)

Daily team

!sh "${CLAUDE_SKILL_DIR}/team.sh"

Request: $ARGUMENTS

If the team above is missing or shows an error, run sh team.sh from this skill's folder and use its output.

Modes

Pick exactly one mode from the request. Do no more than that mode asks.

Session

Read prompt.md in this skill's folder, sections "Work the brief" and "Writing rules", and follow them. Output the session itself, with no lead-in. Write each section heading as a level-two heading: ## Pitch, ## Clash, ## Decision or ## Deadlock, ## Build notes.

If the session ends in a Deadlock, apply the "Referee" section of prompt.md once and output its ruling. Never run a second referee.

Build

Also read the "Build an artifact" section of prompt.md. Copy this checklist and tick it off:

Build progress:
- [ ] Session written and reviewed
- [ ] Artifact 1 written, member 1 leading, and reviewed
- [ ] Artifact n written, member n leading, and reviewed (one line per artifact)

Write artifact n to team-builds/<date>/artifact-<n>.html, relative to the working directory the user started in, not this skill's folder or a scratchpad. Write one complete file per artifact, not parts joined later. Member n leads artifact n, so each artifact takes its look and layout from a different member. Give the user the paths at the end.

Review

Check the session and each artifact against this list. If any item fails, fix it and check again. Give the paths only when every item passes.

• The session has ## Pitch, ## Clash, and then ## Decision or ## Deadlock. The Clash has 6 to 10 lines, and no member speaks more than 3 times.
• No word from the banned list in "Writing rules", no exclamation marks, and no emoji.
• The artifact has a title element and a viewport meta tag, and is under 100 KB.
• No external scripts, styles, fonts, images, or iframes, no @import, no fetch or other network calls, and no form that posts to another site.
• The decision's concept is visible on the first screen, and a credits line at the bottom names each member.
• It loads with a worked example already entered.
• Inputs update the result at once, with no submit button for calculations, and labels are 1 to 3 words with no paragraphs of instructions.

Requirements

• Team, session, and build: sh and awk.
• publish and recruit: Claude Code, git, gh signed in, and Node 20 or later. Chrome is optional, for images.

daily-team

A new four-person creative team every day, and the small web tool it built. This repository is the record: each day's team, its plan, and its asset, kept in days/ and listed in the archive.

Everything in days/ is written by a language model and checked by scripts, not by hand. The model and provider for each day are on its page.

Today

2026-10-07: The Ear Card

Task: Stretch sequence timer: the user lists stretches and hold times, and the timer runs through them.

Decision: The page is a 560×400 px card where the user lists stretches, hold times, and a "both sides" mark, then runs the sequence by ear through three tone patterns and vibration, with a shareable link carrying the list.

Open the day · Use it · Archive

This week

Season 2, since 2026-10-05. New this season: Taxidermist, Locksmith, Upholsterer, Glassblower, Calligrapher, Blacksmith, Tailor, Beekeeper, and the first 36 guests. A new season with 8 new roles and 8 new methods is drafted every Monday.

Vote on what gets built next

Suggest a task or vote with a thumbs-up on open suggestions. The most voted are screened every Monday.

No open suggestions yet.

Recent days

All 5 days

How it works

1. The team. team.sh turns the date into four members, each with a role, a method, and a stance, plus a constraint and a task. From season 3, each member also has a temperament: how they argue. The same date always gives the same team. Within a season no role or method repeats within 7 days, no constraint within 22 days, and no task within 61 days.
2. The guest. From season 2, member 4 is a guest who brings one method: a historical figure, a figure from myth, a character from an old book, an archetype, a person from the future, or a creature. Historical guests never speak as the person. Other guests speak in character, in original words, never quoting the source. Real people and authors must have died at least 100 years ago, myths are used in their traditional form, and gods of living religions, brands, and modern franchises are excluded.
3. The cast. Every Monday a language model drafts the next season: 8 new roles, 8 new methods, 4 new temperaments, and 2 new guests join, the longest-serving leave, and voted suggestions become tasks. Scripts check the draft and it opens as a pull request for review. Seasons start on a future Monday, so past days never change.
4. The work. Each morning a language model plays the team. Each member pitches their own version of the task in the voice of their craft. They argue for 6 to 10 lines, then the lead decides. If the team splits, a neutral referee picks one side. Then they build their version as one HTML file.
5. The checks. The file must have a title and a mobile layout, stay under 100 KB, and make no network requests. A file that fails is retried once, then marked "needs review".
6. The record. The workflow takes a screenshot, draws the team card, writes the day page, updates the archive and this README, and commits the result. Days that fail are recorded too.
7. The weekly report. Each Sunday a script gathers the week's changes, days, seasons, and GitHub activity, and a language model writes a short summary from those facts only. A check rejects any number or name that is not in the facts. Reports are in reports/.

Take part

Suggest a task, vote with a thumbs-up on open suggestions, or read CONTRIBUTING.md.

Run your own

Fork the repo, then add one secret:

• OPENROUTER_API_KEY for OpenRouter's free models (the default). Create the key with a credit limit.
• Or CLAUDE_CODE_OAUTH_TOKEN from claude setup-token to use a Claude subscription, and set the repository variable PROVIDER to claude. Each day uses 2 to 5 requests from your plan's limits.

Optional variables: MODEL, CLAUDE_MODEL (default claude-opus-5-5), ARTIFACTS (0 to 4, default 1), TZ, and PAGES (true to publish the tools on GitHub Pages). To record a day by hand with your Claude login, run sh daily.sh. It commits and pushes.

The folder also works as a Claude Code skill: clone it into ~/.claude/skills/daily-team, then use /daily-team.

Safety

• Keys are passed only to the step that calls the model. They are never written to disk or printed.
• The Claude provider runs with no tools, no MCP servers, no user settings or hooks, and no CLAUDE.md, so the model can only return text.
• Every asset gets a Content-Security-Policy that blocks network requests. On GitHub Pages, assets run only inside a sandboxed frame.
• Free models can log prompts. Do not put private material in brief.md.
• See SECURITY.md to report a problem.

License

Code and content: MIT.

Design: the Sam Creates design system, in brand/. The fonts are Mozilla Headline and Mozilla Text under the SIL Open Font License 1.1; see brand/fonts.