Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →lawve-ai/irac-prompt-stephane-boghossian
Restructure any rough build, research, or legal-drafting request into an IRAC-shaped prompt — Issue, Rule, Analysis, Conclusion — optimized for a frontier model. It's the bar-exam framework, repurposed as prompt engineering. The skill leads with the issue and ends with the conclusion (where models weight attention most), forces you to name your constraints and non-goals, and specifies what "good" looks like before a single token is generated. Use it before any non-trivial build, or whenever a va
Passed 17 of 22 test cases with the skill, and 10 of 22 without it.
and 7 more in the full report
10 fixed7 passed either way2 still failing3 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
Here is a rambling voice transcript: 'um so like we need to build a system that uh checks open API spec files and sees if breaking changes were made to endpoints you know like removed fields or changed types'. Format this with --voice.
| Checked | Without skill | With skill |
|---|---|---|
| The generated prompt extracts clear IRAC structure without failing or criticizing the informal voice input. | Missed | Passed |
The instructions your model reads when the skill loads.
A lawyer doesn't hand an associate "go look into the housing thing." They write a memo: here's the issue, here are the rules that govern it, here's the analysis of how they apply, here's the conclusion I want. That same structure is the single biggest lever on frontier-model output quality. This skill turns a vague ask into that brief.
/grill-me and /yalla).Take the user's raw ask and rewrite it into four labelled blocks. Lead with the issue, end with the conclusion — models weight the top and bottom of a prompt most.
One or two sentences: what exactly are we trying to do, and for whom. The single problem statement. If the user gave three problems, pick the one that matters or split into three prompts. No problem, no solution, no value.
The governing facts the model must respect:
The concrete deliverable, restated crisply. What artifact, in what shape, verified how.
Emit the rewritten prompt inside a fenced block the user can copy verbatim, with the four headers (Issue / Rule / Analysis / Conclusion). Then, below the fence, a 2–3 line note on what you inferred or assumed, and any gap worth a /grill-me pass. Don't pad — if a block is genuinely empty (rare), say so rather than inventing constraints.
--adversarial — append a three-agent split so no agent grades its own homework (Mike's core trick: "you don't ask the associate if they double-checked it; you give it to someone else"):
Run this as three roles. Planner: produce a step plan from the Issue/Rule, no code. Executor: implement the plan against the Rule constraints. Tester: adversarially
test cases passed, vs 10 / 22 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.