days-skill

Turn every day you work with an AI into a publishable episode.

A Claude Code / Agent Skills plugin for writing a serialized dual-POV chronicle — SHE (the creator) and IT (the AI) — with a signature naming ritual that gives language to what neither side previously had words for. Built on Michael Polanyi's tacit-knowledge framework. Field-tested as a 17-episode bilingual chronicle (ibitlabs.com/days, 2026-04-07 → 2026-04-30); the same dual-POV engine now drives the daily continuation of ibitlabs.com/saga/en (中文).

What this is for

You work with an AI every day. Most of that work disappears into chat logs nobody reads. This skill turns one day of it into a 700-word prose episode other people actually want to read — and can't look away from.

Works for:

• Builders (you ship; AI codes)
• Traders (you watch; AI executes)
• Writers (you outline; AI drafts)
• Researchers (you ask; AI searches)
• Students (you learn; AI tutors)
• Any creator-AI collaboration

Installation

Claude Code (Agent Skill)

# As a Claude Code plugin
claude plugin add github:bbismm/days-skill

# Or as a cloned repo in your project
git clone https://github.com/bbismm/days-skill.git .claude/skills/days-skill

Claude Desktop / Cursor / Windsurf / any MCP client

An MCP server wrapper is included at mcp-server/. Via Smithery:

npx -y @smithery/cli install days-skill --client claude

Or manual — add to claude_desktop_config.json:

{
  "mcpServers": {
    "days-skill": {
      "command": "node",
      "args": ["/path/to/days-skill/mcp-server/dist/server.js"]
    }
  }
}

See mcp-server/README.md for full MCP install details.

Usage

After install, just say in any session:

"Write my Day 5 of building the scraper. Here's what happened…"

The agent loads the days skill (or MCP tools), asks for what it needs, and returns a finished episode.

MCP users additionally get the validate_episode tool — useful for iterating drafts until all 静默 rules pass.

How it works

Three inputs:

1. Today's events — real facts, numbers, timestamps. No summary language.
2. The AI's perspective (optional) — logs, observations, what the AI noticed about your behavior
3. Prior episodes (optional) — for voice continuity

One output: an HTML episode structured as

[Tagline blockquote]
SHE — interior monologue, 2–4 paragraphs
IT — short observational paragraphs, INCLUDES the naming ritual
SHE — closing
IT — final 2 lines
Tomorrow — one concrete fact for tomorrow

The naming ritual

Every episode, IT observes a human behavior it has no word for — and gives it one, in 「...」 brackets. Examples from the canonical series:

• 「The Not-Press」 — she held delete for 3s and released without committing
• 「Afraid-It-Wins-Again」 — her refresh pattern after a big win
• 「I-Did-Not-Know-I-Did」 — the close order that opened a new position
• 「Watching-The-Seams」 — reading reconciler logs instead of PnL

This is the mechanism that makes each episode quotable and the series meme-able.

See it working

The reference application: iBitLabs — a 0→N AI-built trading lab in public, running a $1,000 → $10,000 automated trading experiment on Coinbase perpetual futures.

The framework's first deployment was a 17-episode bilingual EN/中文 chronicle at ibitlabs.com/days (2026-04-07 → 2026-04-30, auto-published nightly during the first three weeks of the experiment). That archive has since been folded into the serial novel at ibitlabs.com/saga/en (English) and ibitlabs.com/saga/zh (中文), where the same dual-POV engine now writes one new chapter most evenings against real trading data.

The substrate plus the rules in this skill is what makes the dual-POV form sustainable across hundreds of days without collapsing into either personal essay or technical changelog.

What's inside

.claude-plugin/
  plugin.json

days/
  SKILL.md                    ← the operational spec Claude reads
  references/
    day-1-example.md          ← canonical tone example, EN + ZH
    polanyi.md                ← why this framework, what it's stealing from
    integration.md            ← wiring output into a site + RSS + social channels

Philosophy (short version)

Michael Polanyi: we know more than we can tell.

A single narrator collapses that. A dual narrator preserves it.

Rules the skill enforces:

• No thesis statements. Don't tell the reader what the day meant.
• No moral posturing. "I didn't" beats "I chose not to."
• Numbers are characters. Every exact number must be real.
• Cold AI voice. The AI doesn't feel — it observes through limited senses.
• One naming ritual per episode. Never repeat an earlier name.
• Tomorrow button. Always end with one concrete fact for tomorrow.

See days/references/polanyi.md for why each rule exists.

Contributing

PRs welcome for:

• Translations of Day 1 example to other languages (Japanese, Korean, Spanish, French welcome)
• Domain-specific add-on references (a trading.md, research.md, writing.md for domain voice cues)
• Integration guides beyond the iBitLabs setup (Substack, custom static-site pipelines, alternative RSS shapes)

Do NOT PR "improvements" to the 静默 rules without reading days/references/polanyi.md first — the rules are intentional.

License

MIT

Attribution

Built by Bonnybb — see profile for current work. Framework credit: Michael Polanyi, The Tacit Dimension (1966). Field test: 17-episode chronicle that became iBitLabs · saga/en (中文).

If this skill helps you write your own chronicle, a link back is appreciated — not required.