skill
food-chain-ideation
Claude Code skill that stress-tests startup ideas with adversarial AI agents — 68 animals, elimination rounds, blind scoring. Your idea either survives or you get 3 pivots
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/CodedRichy/food-chain-ideation itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 17, 2026Food Chain
Stress-test your startup idea before you build it. A Claude Code skill. 68 AI agents attack your product idea from every angle in live elimination rounds. What survives is worth building. What doesn't gets three pivots.
What happens when you run it
You describe your idea. The God Agent reads it, selects animals matched to your specific threat vectors, states its kill hypothesis, and asks how many agents you want.
Then this happens:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
ROUND 1 · 5 animals remaining
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🪱 TAPEWORM — Instagram API dependency
Instagram Graph API rate-limits at 200 calls/hr. TikTok Content Posting API
requires individual app review with revocable access. Your entire product
sits on rented land with a landlord who's building the same house.
Kill Shot: The platform adds your feature and terminates the relationship.
🐦⬛ CROW — Timing assumption forensics
Platform algorithms are independent — TikTok peaks over 72hrs, YouTube Shorts
over 7 days. Cross-platform "catch the wave" is meaningless. Speed provides
zero algorithmic advantage on destination platforms.
Kill Shot: You're selling a stopwatch in a race that isn't timed.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🐦⬛ Crow consumes 🐢 Tortoise
Absorbed: authenticity/shadowban concern
Idea patch: Pivot from "publish automatically" to "draft instantly, creator approves"
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Four rounds later, the original "viral content repurposer" was dead. The surviving idea was a cross-creator structural intelligence platform. It shared almost no DNA with what went in. The Crow's final kill: "Structural intelligence produces convergence. Thousands of creators using the same tool produce structurally identical posts. The better the product works, the faster it poisons its own signal."
No critique tool has that insight sitting in it. It emerged because each attacker was isolated, role-locked, and forced to find a new angle after each patch.
Full transcript:
battles/viral-content-repurposer.md
Three outcomes
A tool that only kills ideas is a pessimist. Food Chain produces all three:
| Outcome | What happened |
|---|---|
| Survived | Developer tool. Early termination applied, because the attacks couldn't land structural kills. Core thesis held. |
| Restructured | B2B SaaS. ICP corrected from "Indian freelancers" to "newly-registered, enterprise-triggered, 0–2 years compliance history." Distribution rebuilt. Moat found. |
| Killed | Office meal coordinator. The stated ICP overwhelmingly uses contracted caterers. The problem the product solved didn't exist. Redirected to B2B caterer procurement intelligence, a completely different business. |
Try it now
claude plugin marketplace add CodedRichy/food-chain-ideation
claude plugin install food-chain-ideation@food-chain-ideation
Then type: food chain my idea: [describe your product]
The animals
68 animals across three libraries. Each has hardcoded behavioral DNA. The God Agent selects, never invents.
🐦⬛ Crow. Forensic. Finds the structural assumption three levels deep that everyone overlooked. Quiet, then devastating.
🪱 Tapeworm. Lives inside the platform you depend on. Invisible until you scale. Then the host adds your feature and terminates the relationship.
🐢 Tortoise. Has seen every wave of disruption since 2009. Has a spreadsheet that works fine.
🦈 Shark. The incumbent with 200k users who can ship your feature as a Tuesday sprint toggle.
🐘 Elephant. Microsoft/Google adds it to the bundle. Free. Tomorrow.
🦂 Scorpion. Compliance. GDPR. SOC 2. The thing you forgot until a customer's legal team sends the questionnaire.
35 product animals. 21 code-specific animals. 12 pitch animals. Full libraries in
animal-library.md,
code-animal-library.md, and
pitch-animal-library.md.
Five skills, one install
| Skill | Purpose |
|---|---|
| food-chain-ideation | Stress-test product ideas. 35 animals, elimination rounds, blind scoring |
| food-chain-code | Stress-test architecture decisions before writing code. 21 code animals |
| food-chain-pitch | Stress-test investor pitches. 12 pitch animals, investor simulation panel |
| apex-to-action | Turn battle output into a 90-day execution plan with validation gates |
| food-chain-monitor | Re-test after pivots or market shifts. 3 animals, 2 rounds, fast verdict |
Stress-test → Pitch → Execute → Monitor → Re-test.
Why this is different
Standard adversarial critique gives you four attacks and a summary. You read them, nod, and build the same thing you were going to build.
Food Chain gives you four attacks where the weakest dies, its insight transfers to the winner, the idea patches itself, and the next round's agents attack the patched version. The attacks compound. The idea evolves. The final surviving argument has been pressure-tested by everything the ecosystem could generate.
Subagent mode. In Claude Code, each animal is a separate agent with zero shared context. Role-lock is architectural. A Crow literally cannot see what the Tapeworm said.
Blind scoring. An independent agent scores the attacks and receives them anonymized as "Attacker A", "Attacker B". No animal names. No God Agent self-assessment bias.
Pivot engine. When an idea is killed, three pivots auto-generate from the wreckage, each mini-battled and ranked. You don't leave with "don't build this." You leave with "build this instead."
Audience agent. After the battle, a simulation of your actual ICP reacts to the evolved idea. "Would you pay for this?"
Usage
food chain my idea: [your idea]
what kills this: [your product]
food chain code: [your architecture decision]
food chain pitch: [your pitch or deck summary]
apex to action: [paste your battle log]
food chain monitor: [paste battle log + what changed]
Install
Via Claude Code plugin system (recommended):
claude plugin marketplace add CodedRichy/food-chain-ideation
claude plugin install food-chain-ideation@food-chain-ideation
Manual install:
git clone https://github.com/CodedRichy/food-chain-ideation.git
# Project-level (this project only):
cp -r food-chain-ideation/skills .claude/skills
# Or global (all projects):
cp -r food-chain-ideation/skills ~/.claude/skills
On Windows, ~/.claude/ is C:\Users\<you>\.claude\.
Self-contained markdown. No build step, no config, no dependencies.
Works in Claude.ai, Claude Code, Cursor, Windsurf, Copilot.
skills/
├── food-chain-ideation/
│ ├── SKILL.md ← product idea battles
│ └── references/
│ └── animal-library.md ← 32 animals, behavioral DNA
├── food-chain-code/
│ ├── SKILL.md ← architecture decision battles
│ └── references/
│ └── code-animal-library.md ← 20 code-specific animals
├── food-chain-pitch/
│ ├── SKILL.md ← investor pitch battles
│ └── references/
│ └── pitch-animal-library.md ← 12 pitch-specific animals
├── apex-to-action/
│ └── SKILL.md ← battle output → 90-day plan
└── food-chain-monitor/
└── SKILL.md ← re-test after changes
Contributing
See CONTRIBUTING.md. Every new animal added by the community makes every existing installation better retroactively.
v2.1.0
- food-chain-pitch. Investor pitch stress-tester with 12 pitch-specific animals
- Real-time data mode. Agents search for current market data when web tools available
- Kill shot confidence tags. CONFIRMED / PROBABLE / SPECULATIVE on every attack
- Scoring reasoning. 1-line justification per scoring dimension
- Audience panel. 3 ICP simulations replace single audience agent
- Founder defense agent. Active defense between attacks and scoring
- Animal selection reasoning. Transparent selection logic in ecosystem design
- Battle gallery on website. 4 real battle transcripts browseable
- Spec alignment. All skills use when_to_use + argument-hint frontmatter
v2.0.0
- Subagent mode. Architecturally enforced role-lock via independent agents
- Blind scoring agent. Anonymous attack evaluation removes self-assessment bias
- Subagent evolution. One-line patch context enables angle evolution without breaking isolation
- Audience agent. ICP simulation reacts to evolved idea after battle
- Pivot engine. Auto-generates and mini-battles 3 pivots when idea is killed
- food-chain-code. Architecture decision stress-tester with 20 code-specific animals
- apex-to-action. Battle output → 90-day execution plan with validation gates
- food-chain-monitor. Delta re-testing after changes with 3 animals, 2 rounds max
- Restructured as multi-skill suite. One install, four skills
v1.3.0
- Input sharpening. Mandatory ICP/geography/problem check before ecosystem design
- Battle quality score. Role-lock integrity, animal specificity, kill shot lethality
- Next-move block. Validate, build first, never build
- Battle log. Compressed session memory for cross-session continuity
v1.2.0
- Pre-flight checks. Ecosystem validated before battle starts
- God Agent hypothesis. Kill prediction stated before round 1
- Early termination rule. Strong ideas declared early
v1.1.0
- Anti-pattern combinations added to animal library
v1.0.0
- Initial release. 16 animals, elimination and absorption loop