Learning HubACFS Academy
Progress
0%
0 of 2020 remaining
  • 1
    Welcome & Overview5 min
  • 2
    Linux Navigation8 min
  • 3
    SSH & Persistence6 min
  • 4
    tmux Basics7 min
  • 5
    Git Essentials10 min
  • 6
    GitHub CLI8 min
  • 7
    Agent Commands10 min
  • 8
    NTM Command Center8 min
  • 9
    NTM Prompt Palette6 min
  • 10
    The Flywheel Loop10 min
  • 11
    Keeping Updated4 min
  • 12
    UBS: Code Quality Guardrails8 min
  • 13
    Agent Mail Coordination10 min
  • 14
    CASS: Learning from History8 min
  • 15
    The Memory System8 min
  • 16
    Beads: Issue Tracking8 min
  • 17
    Safety Tools: SLB & CAAM6 min
  • 18
    The Art of Agent Direction12 min
  • 19
    Case Study: cass-memory15 min
  • 20
    Case Study: SLB12 min
    NOW
Back to Home
Back
20/20
Lesson 20
12 min

Case Study: SLB

From tweet to working tool in one evening

New to ACFS?

Complete the setup wizard first to get the most from these lessons.

Go to Choose Your OS
Goal

See how a tweet becomes working code in one evening: 76 beads, 268 commits, from idea to ~70% complete in hours.

The Spark: From Tweet to Tool

On December 13, 2025, a conversation on X about AI agents accidentally deleting Kubernetes nodes sparked an idea: what if dangerous commands required peer review from another agent?

The WarGames Insight

"You know how in movies like WarGames they show how the two guys have to turn the keys at the same time to arm the nuclear warheads? I want to make something like that where for potentially damaging commands, the agents have to get one other agent to agree with their reasoning and sign off on the command."

Two-person rule for AI agents

The idea was simple: like the "two-person rule" for nuclear launch codes, agents should need a second opinion before running destructive commands like rm -rf, kubectl delete, or DROP TABLE.

Immediate Action: The First Hour

Instead of just noting the idea for later, the flywheel approach is to start immediately while the idea is fresh.

December 13, 2025 Timeline

3:55 PMIdea sparked from tweet
~4:30 PMInitial plan drafted with Claude Code
5:25 PMPlan document published
EveningMulti-model feedback gathered
NightBeads created, implementation started
Pro Tip
The key insight: an initial plan within the first hour, even if rough, is worth more than a perfect plan days later. The agents will help refine it.

The Feedback Loop: Four Models, One Plan

Once the initial plan existed, it was sent to multiple frontier models for review and improvement:

Claude Opus 4.5

Architecture refinement

Gemini 3 Deep Think

Edge case analysis

GPT 5.2 Pro

Security considerations

Claude (synthesis)

Combining all feedback

The feedback was then integrated by Claude Code, with multiple verification passes to ensure nothing was missed:

bash
1# First pass: Integrate all feedback
2cc "Revise the plan document using all the feedback.
3Make sure ALL changes are reflected properly."
4
5# Second pass: Verification
6cc "Go over everything again. Did we miss anything?"
7# Result: Found small oversights
8
9# Third pass: Final check
10cc "One more careful review. Any remaining gaps?"
11# Result: Found 2 more edge cases
Note
Each verification pass found something. This is why multiple passes are critical - they catch problems in the planning phase when they're easiest to fix.

Plan to Beads: Making It Executable

The refined plan was then transformed into structured, trackable beads. The prompt was carefully crafted to ensure thoroughness:

bash
1cc "First read ALL of the AGENTS.md file and
2PLAN_TO_MAKE_SLB.md file super carefully.
3Understand ALL of both! Use ultrathink.
4
5Take ALL of that and elaborate on it more, then create
6a comprehensive and granular set of beads with:
7- Tasks and subtasks
8- Dependency structure
9- Detailed comments making everything self-contained
10- Background, reasoning, justification
11- Anything our 'future self' would need to know
12
13Use the bd tool repeatedly to create the actual beads."

Final Beads Structure

14
Epics
62
Tasks
76
Total Beads

Smaller than cass-memory's 693 beads, but still comprehensive enough to capture the full implementation.

Then, just like the plan itself, the beads went through verification passes:

bash
1# Beads verification prompt
2cc "Check over each bead super carefully:
3- Does it make sense?
4- Is it optimal?
5- Could we change anything to make the system work better?
6
7If so, revise the beads. It's a lot easier and faster
8to operate in 'plan space' before implementing!"

What SLB Does

The Simultaneous Launch Button implements a two-person rule for AI coding agents:

CRITICAL2+ approvals

System destruction, database drops

DANGEROUS1 approval

rm -rf, git push --force

CAUTIONAuto (30s)

Single file delete, branch remove

SAFESkip

Temp file cleanup, cache clear

  • Client-side execution: Commands run in the user's shell, inheriting all credentials
  • Command hash binding: Approvals tied to exact commands via SHA-256
  • Pre-flight validation: Automatic dry-runs for supported commands
  • Rollback capture: System state saved before dangerous operations
  • Agent Mail integration: Reviewers notified automatically

The Implementation Sprint

With beads ready, the agent swarm began implementation. The project was smaller than cass-memory, but the workflow was identical:

bash
1# Launch agents
2ntm spawn slb --cc=3 --cod=2
3
4# Each agent runs:
5bv --robot-triage # What's ready?
6bd update <id> --status in_progress
7# ... implement ...
8bd close <id>
9
10# Commit agent runs every 15-20 min
11cc "Commit all changes in logical groupings with
12detailed messages. Don't edit code. Push when done."

Implementation Results

268
Total Commits
Go 1.21+
Built In
~70%
Day 1 Complete

By dinner time, about two-thirds of the project was complete. The agent swarm continued working while the developer ate, pushing commits autonomously.

Small vs Large Projects

Compared to the 693-bead cass-memory project, SLB's 76 beads allowed for some workflow optimizations:

Small Project (SLB)

  • 76 beads (14 epics, 62 tasks)
  • 3-5 agents sufficient
  • Faster verification passes
  • Easier to track in bv
  • One evening to ~70%

Large Project (cass-memory)

  • 693 beads (14 epics, 350+ tasks)
  • 10+ agents needed
  • Multiple planning sessions
  • Graph analysis critical
  • One day to ~85%
Pro Tip
Start with smaller projects to learn the workflow. Once you're comfortable with 50-100 beads, scale up to larger projects.

Lessons Learned

1

Act immediately on good ideas

An hour from idea to initial plan keeps momentum high

2

Multi-model feedback finds blind spots

Each model brings different perspectives and catches different issues

3

Multiple verification passes are essential

Each pass found something - never skip this step

4

Smaller projects are great for learning

76 beads is manageable while still demonstrating the full workflow

5

Document everything

The conversation transcripts become valuable learning resources

Try It Yourself: Weekend Project

Pick a small tool idea (something that would take you a day or two manually) and try this workflow:

bash
1# Hour 1: Draft initial plan
2cc "I want to build [your idea]. Help me create a
3detailed plan document covering architecture,
4features, and implementation approach."
5
6# Hour 2: Multi-model feedback
7# Send plan to 2-3 different frontier models
8# Collect their suggestions and improvements
9
10# Hour 3: Synthesize and create beads
11cc "Read the plan and all feedback. Create a
12revised plan incorporating the best suggestions."
13
14cc "Convert the plan into 50-100 beads with
15dependencies. Use bd CLI."
16
17# Hour 4+: Implementation
18ntm spawn myproject --cc=2 --cod=1
19# Let the swarm work!
20
21# Every 15-20 min: Commit agent
22cc "Commit all changes with detailed messages."
Note
For your first flywheel project, aim for something with 50-100 beads. CLI tools, utilities, and small libraries are perfect candidates.

Ready to level up?

Mark complete to track your learning progress.

Previous
Case Study: cass-memory