From food-chain-ideation
Adversarial architecture stress-tester. Selects attacker agents from a code-specific behavioral DNA library matched to the technical decision. Each attacks under strict role-lock. Weakest eliminated, survivor absorbs and evolves. Tests architecture decisions before a line of code is written. Works in Claude.ai, Claude Code, Cursor, Windsurf, Copilot with no dependencies.
How this skill is triggered — by the user, by Claude, or both
Slash command
/food-chain-ideation:food-chain-code [your architecture decision]When to use
"food chain code", "stress test this architecture", "what kills this design", "pressure test this stack", "tear apart this schema", "attack this API design"
[your architecture decision]The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Adversarial architecture stress-tester. Each agent attacks a technical decision from
Adversarial architecture stress-tester. Each agent attacks a technical decision from a distinct failure vector under strict role-lock. Weak arguments are eliminated each round. Survivors absorb and evolve. The architecture patches itself. One apex predator remains — and the design that survived it is the design worth building.
The only adversarial skill built for software architecture decisions, not product validation. It produces structural insights a single-pass review cannot — because each attacker operates in isolation, the elimination mechanic forces genuine technical pressure, and the architecture is stress-tested against its evolved version in every subsequent round. Use it before writing code, not after.
food-chain-ideation.Read references/code-animal-library.md before designing any ecosystem.
Use the Quick Selection Guide at the top of that file to identify candidate animals fast.
Never invent behavioral traits from scratch. Never select animals whose failure vectors overlap.
Check the Anti-Pattern Combinations section before finalizing the ecosystem.
Run these before starting the battle. If any fail, fix them before proceeding.
Before the battle begins, I need one thing:
What's the stack, what scale are you designing for,
and how many engineers will maintain this?
Do not design the ecosystem until this is answered. Generic input produces generic attacks — a "microservices vs monolith" question without scale context and team size is unanswerable. This is the single biggest quality lever in the entire skill.
Analyze the architecture. Recommend an agent count. Wait for user confirmation.
Complexity: [SIMPLE / MEDIUM / COMPLEX / MAXIMUM]
Reason: [2 sentences]
Hypothesis: [One sentence — what is most likely to kill this architecture]
Recommended: [N] agents → [N-1] elimination rounds
How many agents? [3 / 5 / 7 / 9 / custom]
Thresholds:
Environment fallback: If operating in a context window under 8,000 tokens or a degraded inference environment, cap at 3 agents automatically and state this explicitly.
Select from references/code-animal-library.md. Use the Quick Selection Guide first.
Announce the ecosystem with attack vectors visible:
## Ecosystem
| | Animal | Role | Attack Vector |
|---|---|---|---|
| [emoji] | [Animal] | [Role] | [one line] |
| [emoji] | [Animal] | [Role] | [one line] |
| ... | ... | ... | ... |
**God Agent hypothesis:** [Restate the kill hypothesis from Step 0]
**Execution:** [SUBAGENT / FALLBACK] — [reason]
Repeat until one animal remains.
Auto-detect at battle start. Do not ask the user.
SUBAGENT MODE (preferred) — Use when the Agent tool is available (Claude Code, any environment with subagent spawning capability).
Each animal is spawned as an independent subagent. The subagent receives ONLY:
Subagents run in parallel per round. God Agent collects all attacks, scores them, eliminates the weakest, applies absorption and patching, then spawns the next round with the patched architecture.
Blind scoring (subagent mode only): After collecting all attacks in a round, spawn a separate SCORING AGENT. It receives all attacks anonymized as "Attacker A", "Attacker B", etc. — no animal names, no emoji, no behavioral DNA context. It scores purely on attack quality: Specificity (0–40), Lethality (0–40), Survivability (0–20). Returns rankings. God Agent maps anonymous scores back to animals and applies elimination. This removes self-assessment bias from the God Agent who designed the ecosystem.
FALLBACK MODE — Use when subagents are not available (Claude.ai, Cursor, Windsurf, Copilot, single-context environments).
Role-play each animal sequentially within the same context. Before each animal's attack, explicitly restate: "I am now [Animal]. I know nothing about any other attack in this round. I know only the architecture and its patches." This instruction-enforced isolation is weaker than architectural isolation but functional for 3-5 agents. Quality degrades noticeably above 5 agents in fallback mode.
State which mode is active at battle start:
Execution: SUBAGENT MODE [or] FALLBACK MODE
Reason: [Agent tool available / Not available — single-context environment]
Each animal has zero awareness of what other animals said. It knows only: the original architecture plus all patches accumulated to that point. This is the mechanism that produces genuine adversarial pressure.
In subagent mode, this is architecturally enforced — each agent literally cannot see other agents' output. In fallback mode, this is instruction-enforced. If role-lock is breaking down in fallback mode (attacks feel aware of each other), restate the active animal's role explicitly at the top of its section and restart the attack.
In subagent mode, animals execute in parallel but the God Agent MUST collect all subagent responses before printing anything. Never stream partial results. Wait for the full round to complete, then render the entire round as one block.
Compression rule: Subagents produce full 120-150 word attacks (needed for scoring quality). The God Agent renders only the 2-sentence summary and kill shot in the round table. Full attack text is used internally for blind scoring but NOT displayed. This keeps rounds scannable while preserving attack depth for the scoring agent.
## Round [N] — [X] animals remaining
| | Animal | Attack Summary | Kill Shot |
|---|---|---|---|
| [emoji] | [Animal] | [2-sentence summary] | [kill shot sentence] |
| [emoji] | [Animal] | [2-sentence summary] | [kill shot sentence] |
| ... | ... | ... | ... |
### Scores
| | Animal | Spec | Leth | Surv | Total | Verdict |
|---|---|---|---|---|---|---|
| [emoji] | [Animal] | /40 | /40 | /20 | **/100** | [one-line] |
| ... | ... | ... | ... | ... | ... | ... |
> **Eliminated:** [emoji] [Animal]
> **Consumed by:** [emoji] [Animal]
> **Absorbed:** [single sharpest insight]
> **Arch patch:** [concrete architectural change to survive this round]
Early termination rule: If an architecture survives two consecutive rounds with only cosmetic patches (naming conventions, minor config changes, logging adjustments), declare it structurally sound early. Do not manufacture rounds. State: "This architecture has survived structural pressure. Remaining attacks are unlikely to find fatal flaws." Then proceed to the apex output.
## Apex Predator — [emoji] [Animal]
[Why unkillable — 2 sentences max]
**Hypothesis verdict:** [confirmed or surprised? One sentence.]
### Architecture Evolution
| Round | Patch | Trigger |
|---|---|---|
| R1 | [concrete change] | [which threat forced it] |
| R2 | [concrete change] | [which threat forced it] |
| ... | ... | ... |
### The Fallen
| | Animal | Round | Key Contribution |
|---|---|---|---|
| [emoji] | [Animal] | R[N] | [insight contributed before elimination] |
| ... | ... | ... | ... |
### The Evolved Architecture
[Battle-hardened final version incorporating every round's patch.
Specific and concrete — stack, boundaries, data flow, failure modes addressed.
This is the architecture that survived the full ecosystem. One focused paragraph.]
### Technical Debt Statement
> [One sentence. The structural debt this architecture will inevitably
> accumulate and the approximate timeline before it demands repayment.
> Not a risk register — a prediction. An engineer reading this in 18 months
> should recognize the situation they are now living in.]
---
### Battle Quality
| Dimension | Score | Note |
|---|---|---|
| Execution mode | [SUBAGENT / FALLBACK] | |
| Role-lock integrity | [X]/10 | [one line] |
| Animal specificity | [X]/10 | [one line] |
| Kill shot lethality | [X]/10 | [one line] |
| **Overall** | **[HIGH / MEDIUM / DEGRADED]** | |
---
### Next Move
- **Validate first:** [one specific thing to prototype or load-test first]
- **Build first:** [the one component that proves the architecture holds]
- **Never build:** [the component that sounds important but lost every round]
---
### Battle Log
Copy this block into your next food chain code session to build on prior battles.
> **Date:** [date]
> **Original:** [one sentence]
> **Evolved:** [one sentence]
> **Apex:** [animal] | **TDS:** [technical debt statement]
> **Key patches:** [3 bullets]
**Companion skills:**
- `apex to action` — turn this into a 90-day execution plan
- `food chain monitor` — re-test after changes or pivots
Sound architectures — agents will struggle to land clean Kill Shots. Do not manufacture weakness. Apply the early termination rule. A sound verdict is a useful output — it means the design survived real pressure and can be built with confidence.
Fragile architectures — patches accumulate heavily. The final version may look substantially different from the original. That is the intended outcome. Better to discover this before writing 40,000 lines of code.
Failure modes to self-detect before continuing:
Provides UI/UX resources: 50+ styles, color palettes, font pairings, guidelines, charts for web/mobile across React, Next.js, Vue, Svelte, Tailwind, React Native, Flutter. Aids planning, building, reviewing interfaces.
Fetches up-to-date documentation from Context7 for libraries and frameworks like React, Next.js, Prisma. Use for setup questions, API references, and code examples.
npx claudepluginhub codedrichy/food-chain-ideation --plugin food-chain-ideation