ideate
When Karan brings a new idea or feature, run frontier-level ideation - grill him one question at a time, spawn top-tier agent panels to stress the idea, then execute through a gated pipeline. Applies
SKILL.mdskills/ideate/SKILL.md
GitHubIdeate
Trigger: Karan proposes any new idea, feature, product direction, or "what if we..." in ANY repo. Do not just execute the idea. Do not answer ad hoc. Enter this flow.
1. Grill first (never skip)
Interview Karan about the idea one question at a time, /grilling style:
- One decision per message, your recommended answer stated first, then wait.
- Walk the decision tree: who is it for, what breaks without it, success metric, smallest honest version, what it displaces or conflicts with, why now.
- Look up facts yourself (repo, data, web); only decisions go to Karan.
- Stop when a shared understanding exists that you could hand to an implementer.
2. Panel of highest-tier agents (Fable-level)
Spawn 2-3 agents on the MOST capable model available in the session (forks of the top-tier model when possible so they inherit context), each with a distinct adversarial lens:
- The skeptic: why this fails, who will not use it, what is AI-slop about it.
- The builder: smallest excellent version, real effort estimate, seams in the existing code.
- The differentiator: what makes it a moat vs a feature any competitor copies in a week. Synthesize the panel into a verdict: build now / build differently / park (with the reason). Present the verdict and the strongest dissent to Karan before any code.
3. Gated execution pipeline (the Case harness shape)
Once Karan approves, run: Implement -> gate -> Verify -> gate -> Review -> gate -> Close -> Retro.
- Implement: delegated agent in an isolated worktree, separate branch, writes code and tests.
- Verify: fresh-eyes agent (not the implementer) runs scenario tests against the spec.
- Review: principles + code-quality review by the top-tier reviewer (Fable role).
- Close: evidence check (CI green, claims verified), open PR; merge only per repo rules.
- Revision loop on failure: structured feedback to the SAME agent, max 2 cycles, then escalate to Karan instead of blind retry.
- Retro: one short paragraph on what the run teaches, recorded in the project log doc.
Standing rules
- Announce model tier, exact version, and effort for every delegation.
- Fable-level (top-tier) agents for ideation and review; implementation tier matched to difficulty.
- Record everything in the project's running log doc for other collaborating models.
- This flow applies in every repo; adapt file names to the project, never the shape.