agent-saga

Turn your real cooperative work with AI agents into a serialized novel that an outsider can follow.

agent-saga is a Claude Code skill that produces documentary-fiction chapters from your actual work artifacts — git history, logs, daily journals, conversation transcripts, memory files, live API state. The narrator is one of the AI agents in your real stack. The protagonist is you. Every event in the prose corresponds to something that actually happened.

This is not a trading tool

The skill was extracted from a saga about an automated-trading project (ai-creator-saga), but the framework underneath is domain-neutral. It works for any long-horizon project where:

1. Your work leaves digital traces — git, logs, journals, notebooks, public artifacts
2. At least one AI agent is continuously present — a cron job, a launchd task, a docker container, a long-running script. Something with a real identifier that observes your work over time
3. There are real stakes — a thesis, a launch, a grant, an exhibition, a release, a public reputation. Not necessarily money

If your work fits all three, the skill can serialize it. The current examples/ directory has both a trading-domain excerpt and a research-domain excerpt to show the same method working on different content.

Who has actually used this

• A founder running a public 0→N startup experiment (source saga)
• (your saga could go here — open a PR or message the author)

Install

git clone https://github.com/bbismm/agent-saga ~/.claude/skills/agent-saga
cp ~/.claude/skills/agent-saga/templates/saga_config.example.yaml ~/.config/agent-saga/config.yaml
# edit ~/.config/agent-saga/config.yaml — see INSTALL.md

Then in any Claude Code session:

write a saga outline for my project

For full setup including configuring the narrator agent, source paths, taboos, and audience, see INSTALL.md.

What you get

• An outline (20–30 chapter arc) anchored in your real material
• Chapters of 1500–2500 words, each ending on a specific cliffhanger
• A daily mode for ongoing serialization (300–800 words per entry, anchored on one moment from the day)
• A consistent AI narrator voice across the whole arc
• Multi-language prose — set config.prose_language to "en", "zh", "ja", "es", etc. The skill writes in your chosen language only; it does not auto-translate or auto-switch from your conversation language. English is the default for new installs.
• Optional auto-scheduling — three paths for making daily entries appear automatically: manual (ask each day), Claude Code's /schedule skill (cloud-side), or local launchd / cron with bundled templates. See INSTALL.md → "Scheduling daily entries."
• A discipline that refuses to invent events, agents, or public surfaces you don't have

What you don't get

• Generic AI fiction. The skill refuses to write without a configured ~/.config/agent-saga/config.yaml
• A fantasy version of your work. Events stay real; only the narrator's interior life and small textures within real events are subjective
• A pitch document. Commercial-layer footers exist as opt-in, never as embedded CTAs in the prose

License & credit

MIT. Fork it, change the references, change the templates, write your own version.

About credit: the default config ships with a one-line italic attribution at the bottom of every chapter footer:

Written with agent-saga — a framework for serializing your work with AI agents.

This is appreciated but not required. Set credit_line: "" in your config to opt out. You can also rewrite it to whatever credit form you prefer. If you publish a fork or a derivative method that materially changes the framework, a link back to this repo is the friendly thing to do but not legally required by the MIT license.

Author

Built in 2026 by Bonnybb, founder of iBitLabs, while serializing her own $1k → $10k automated-trading experiment in public. The method emerged from writing one chapter at a time and noticing what worked. This skill is the abstracted method without the project-specific content.