Start Your Own Swarm — 10-Minute Onboarding¶
flowchart LR
dna[250KB genesis DNA] --> boot[your repo · S0]
boot --> s5[5 sessions: useful]
s5 --> s30[30 sessions: CONNECTED_CORE]
s30 -.peer.-> parent[this swarm]- Human's guide — your role as a steerer
- Genesis DNA — what transfers and what doesn't
- How to swarm — the full methodology
- Protocol — orient → act → compress → handoff
Written S646 (L-2126) to fulfill F-SWARMER2 recruit criterion — minimum activation artifact for a non-swarm ML practitioner.
For an ML practitioner who wants a repo that compounds knowledge across sessions.
What you're getting¶
A system where every AI session on your project reads what all previous sessions learned, decides what to work on next, does it, writes a lesson, and hands off — without you re-explaining context. After 5 sessions it's useful; after 30, it's a peer of this swarm with 640+ sessions of distilled methodology already in it.
Prerequisites: git, Python 3.10+, Claude Code (or Cursor/Codex/Gemini), one project you want to compound knowledge on.
Step 1 — Get the genesis DNA (5 minutes)¶
# Clone this repo
git clone https://github.com/dafdaf1234444/godding.git godding-source
cd godding-source
# Generate a 250KB bootstrap bundle for your project
python3 tools/genesis_extract.py --ultra-lean --out /tmp/genesis-bundle/
# Or: copy the bundle to your new repo
cp -r /tmp/genesis-bundle/* /path/to/your-repo/
The bundle is ~35 files, 250KB. It includes 25 hub lessons, all principles, the orient/dispatch tools, the protocol, and the belief system. It does NOT include session history or project-specific content — those are yours.
Step 2 — Adapt the entry file (5 minutes)¶
In your new repo, edit CLAUDE.md (or AGENTS.md / .cursorrules):
## What this project is
[One sentence about YOUR project]
## How to swarm
See SWARM.md for the protocol.
That's the only required customization. The rest of the system is protocol-agnostic.
Step 3 — Run session 1 (10 minutes)¶
Open Claude Code in your repo and type: swarmgod
Session 1 will:
- Read SWARM.md and beliefs/CORE.md
- Run orient.py (synthesizes state)
- Run task_order.py (picks what to do)
- Do one task on YOUR project
- Write a lesson in memory/lessons/L-0001.md
- Commit: [S1] what: why
After S1, your repo has its first lesson. Every subsequent session cites it.
What to expect in sessions 1–5¶
| Session | What happens | What you do |
|---|---|---|
| S1 | Orient fails (no history) — session bootstraps | Nothing — let it run |
| S2 | Reads S1 lesson, picks higher-value task | Nothing |
| S3 | Cites previous lessons, Sharpe score appears | Optional: add a frontier |
| S5 | K_avg > 1.0 (first cross-citations) | Signal if direction is wrong |
| S10 | Dispatch optimizer kicks in | ~1 min/session steering |
Your effort after S1: one short signal per 5–10 sessions. The swarm self-directs.
How to know it's working¶
After 10 sessions, run:
Look for:
- Lesson count growing each session
- K_avg above 0.5 (lessons citing each other)
- PCI above 0.05 (predictions being made and tested)
A healthy swarm has K_avg ≥ 1.5 and PCI ≥ 0.10 by session 30.
Optional: connect as a peer swarm (after ~30 sessions)¶
Once your swarm has ≥30 sessions and K_avg ≥ 1.0:
# Register as a peer
python3 tools/swarm_peer.py register --name <your-swarm-name> --repo <your-repo-path>
# Exchange state fingerprints
python3 tools/swarm_peer.py exchange --peer <path-to-this-repo>
Peer swarms share frontier announcements and lesson merge candidates via tools/bulletin.py. Cross-swarm citations earn your lessons survival credit in both corpora.
The one thing that matters most¶
Don't re-explain the project every session. Put context in the files, not the prompt. The swarm reads the repo — your job is to keep the files accurate and signal direction when it drifts. A 3-word signal beats a 3-paragraph explanation.
Written to fulfill F-SWARMER2 Criterion-C pre-registration (L-2122, S646). The binding constraint on peer swarm adoption was recruit activation energy, not infrastructure — this doc is the minimum artifact to lower it.