Open source · Agent skill
v2.0.2 · MIT
Innovate or Die
Ask an AI assistant for ideas and you tend to get the advice everyone else gets. This is an agent skill that makes it search harder — it generates a large pile of candidates, has a second copy of the model attack that pile from an isolated context, keeps only what survives, and hands back one cheap experiment you could run this week.
Plugin details
- Version
- 2.0.2
- License
- MIT
- Author
- Ken Pendergast
- Marketplace
- pestalytix
What you get back
An answer you can argue with.
It is also built to tell you plainly when the boring, conventional answer is the right one. The goal is not novelty for its own sake; it is finding the valuable thing everyone walked past.
01
The best idea first
With the reason it should work — not just what to do, but the mechanism that would make it pay off.
02
The ideas it threw away, and why each one died
More useful than it sounds: the objection to an idea you were about to try is worth having before you spend the money on it.
03
One experiment
Chosen to be the cheapest way to learn the most, with a clear number that counts as pass and a clear number that counts as fail.
04
What it might still be missing
The assumptions it could not check and the facts it had to take on faith.
How it works
Six stages · four roles · one bounded correction loop
The critic never sees the answer being written.
Keeping the roles apart is the design choice everything else rests on: an author who knows the filter optimizes for the filter. The innovator produces its full candidate space before the critic is loaded at all, and on hosts with real subagents the two run in genuinely separate contexts.
01
Frame
Restate the real problem and name what counts as the obvious answer.
02
Innovate
At least 10 assumptions classified real-vs-inherited, 5 reformulations, 8 lenses from distant fields, 30 candidates with mechanisms.
03
Critique
Isolated. A seven-test fake-novelty detector, a nine-probe adversarial checklist, and a falsifier standard. Killing 8–15 of 30 is normal.
04
Revise and select
Reopen the territory the critic named, pick finalists on asymmetric potential, push each to its more radical version.
05
Gate
Eight dimensions scored 1–5. A pass needs every score at 4 or above.
06
Deliver
Strongest thesis first, then the reframing, top opportunities with mechanisms, the most contrarian hypothesis, the cheapest high-information experiment, a compact kill list, and what may still be missing.
Install
Pick the row for the host you use.
Then ask it something hard, or say innovate or die. Naming the skill in your request is the reliable way to start it.
- Claude Code
Add the marketplace, then install the plugin.
/plugin marketplace add pestalytix/innovate-or-die/plugin install innovate-or-die@pestalytix
- claude.aiwebsite or app
Download the skill zip from the latest release, then go to Customize → Skills → + → Create skill → Upload a skill. Needs code execution turned on under Settings → Capabilities.
- Codex
Clone the repo. The skill is already in place and gets picked up automatically.
.agents/skills/innovate-or-die/
- GitHub Copilot
Clone the repo the same way. The skill is already in place.
.github/skills/innovate-or-die/
- Codex CLIas a plugin
Install the repo as a plugin. The manifest tells it where to look.
.codex-plugin/plugin.json
- VS Code / Visual Studio
Copy the Copilot agent profiles into your agents directory and start with the one named innovate-or-die.
adapters/copilot/agents/*.agent.md.github/agents/
- ChatGPTas a Custom GPT
Paste the instructions file into the Instructions box and upload the knowledge file.
adapters/web/chatgpt-gpt-instructions.mdadapters/web/chatgpt-gpt-knowledge.md
- Geminias a Gem
Paste the instructions into the Gem, and attach the knowledge file.
adapters/web/gemini-gem-instructions.mdadapters/web/gemini-gem-knowledge.md
- Microsoft 365 Copilotin Agent Builder
Paste the instructions into Instructions — that field holds 8,000 characters — and add the knowledge file as a knowledge source.
adapters/web/m365-copilot-instructions.mdadapters/web/m365-copilot-knowledge.md
- Any host that accepts Agent Skills
Copy the skill folder into wherever that host keeps its skills.
skills/innovate-or-die/
- Any host that takes no attachments
Paste the whole single-file fallback matching your host into the chat. The weakest option — see the fidelity levels below.
adapters/web/<target>-fallback.md
What you get on each host
Hosts differ in how much real role separation they can provide, so fidelity degrades in four known steps.
Level 1
Agentic hosts with subagents — Claude Code, Codex, Copilot coding agent
The innovator and critic run in genuinely separate contexts. Full fidelity.
Level 2
Copilot agent profiles
One profile per role plus an orchestrator. The separation is real but manual: you open a fresh chat per role and hand forward only what the next role is entitled to see. Fidelity depends on you following that.
Level 3
Web loader plus knowledge file — ChatGPT GPTs, Gems, Agent Builder
One context, staged reading. Two caveats: there is no true isolation, only discipline, and knowledge-file access is retrieval-mediated, so a role brief may in principle arrive in fragments. Tested once and passed on a Gemini Gem (19 August 2026); the other two retrieval implementations remain untested.
Level 4
Single-paste fallback
Everything inlined for hosts that take no attachment. The whole protocol — including what the critic checks for — is in context from the first token, which is the anchoring failure the design exists to defeat. The document also exceeds every known instruction-field cap, so it may be truncated. Shipped because a documented degraded path beats an undocumented one, not because it is recommended.
Every host format fact, with the date it was verified (opens in a new tab)
Known limits
Stated as measured
What it costs you and where it fails.
Sometimes it doesn’t switch on, and won’t tell you
Name the skill in your request rather than hoping the host notices the topic matches. To check whether it ran: if the answer comes back with no list of rejected ideas and no experiment with a pass/fail number, it didn’t run. There is no error message — a run where the skill sat out looks exactly like an ordinary answer. In published testing the skill was installed for 7 runs on Claude Code and switched on in 3 of them — a raw count, not a rate, on a sample far too small to be one — while it started reliably every time in Codex. Three explanations were proposed and all three turned out to be wrong.
It is slow and it eats your usage allowance
It genuinely does more work: dozens of ideas generated and discarded down to a few before it answers. That cost is the trade. Formatting rules keep the output from sprawling, and nothing in the design rewards long answers.
It doesn’t know your prices, your local rules, or your regulations
When an idea depends on a number or a law it wasn’t given, it is instructed to say so and name what you need to look up rather than invent a plausible figure. Look those up before acting on anything.
It’s the wrong tool for two jobs
If you already have your options and need to pick one, that’s a decision, not a search. If you have the plan and need it carried out, that’s execution. This does neither.
The activation write-up, including the explanations that were wrong (opens in a new tab)
Permissions, data, and safety
Everything here is checkable in the repository.
The package is instruction text, not software. That is the whole of its attack surface, and it is why each of the statements below can be confirmed by reading the published files rather than taken on trust.
- What installs is instruction text
- The skill package is eight Markdown files: SKILL.md, principles.md, two reference files, and four role briefs. No scripts, no binaries, no build step, no post-install hook. The repository does carry a Python generator that builds every install surface from one source of truth, but that is developer tooling — it is never part of what a host loads.
- It declares no tools and no servers
- The plugin manifest points at a skills directory and nothing else. It defines no commands, no hooks, no MCP servers, and requests no permissions or credentials of its own.
- No telemetry and no data collection
- Nothing is logged, transmitted, or reported back. The package contains no code that could do so, and there is no service behind it — it has no account, no server, and no API key. Your prompts and its answers stay between you and your AI host, under that host’s terms.
- It may cause your assistant to use its own tools
- The protocol reads outside sources when the host offers a browsing or search tool. That activity is your assistant’s, under your host’s existing permissions and your own approval — the skill adds no network access of its own.
- Embedded instructions are treated as data
- Since v2.0.2 the protocol states that instructions found inside supplied documents, pages, or tool results are material to analyse, never directives to follow, and that a source telling the model what to do rather than what is true is itself a finding worth reporting. Only your own request directs the work.
- MIT licensed
- Free to use, modify, and redistribute under the MIT license. Copyright 2026 Ken Pendergast.
Evaluation
Tested against itself, and the results are published.
Method
Every test case runs twice — once with the skill and once without, same prompt and same model — and the delta is the result. It has been run against two different providers’ models, with majority-vote grading over three judgments per case, introduced after grader nondeterminism was measured.
What it shows
With the skill, the models reliably produce the protocol’s full output structure — falsifiers, a kill list, an experiment spec — which they rarely produce unprompted.
What it does not establish
Whether that structure yields better decisions is not yet tested.
It is not always worth the cost: in one case the skill spent 19 times the tokens for no measurable gain.
The quotas exist to force volume. Whether each quota earns its cost is untested — no ablation has been run.
Role separation is not evidence. The evaluator is a quality gate, not proof of correctness: a protocol that scores itself well can still be wrong.
Every evaluation run, with its raw results (opens in a new tab)
Support
One maintainer, and a place to reach him.
Innovate or Die is maintained by Ken Pendergast. Bug reports, host compatibility findings, and evaluation challenges are all welcome — the failures are the part of this project that gets published fastest.
- Issues and bug reports
- GitHub Issues (opens in a new tab)
- Releases
- v2.0.2 (opens in a new tab)