Examples

April 26, 2026 · View on GitHub

Auto-research, but for decisions instead of knowledge.

Karpathy's autoresearch spends compute to think better about facts — search → read → refine → repeat instead of answering in one shot. Autodecision applies the same loop to decisions: hypothesize → simulate → critique → red-team → refine, until the insights stabilize.

Most AI gives you one answer in one call. Autodecision runs 20-50 calls across 5 independent personas, an anonymized peer review, an adversary, and a sensitivity pass — then a Judge mechanically measures whether the insights have stabilized before writing the brief.

Normal AI vs autodecision

Normal AI	Autodecision
One answer	5 independent worlds argue
Linear reasoning	Iterative refinement until convergence
No critique	Built-in adversary and peer review
Static snapshot	Time-evolving, first- and second-order effects
Confident	Explicit [min, max] probability ranges per effect
Generic	Grounded in web search — real comps, benchmarks, precedents

The real product isn't the council, the simulations, or the brief. It's a system that refuses to accept the first answer.

What makes a good autodecision question?

Autodecision is built for a specific shape of question. Match the shape and it earns its overhead. Mismatch it and a single LLM call is faster.

Good shape

Specific action + timeframe — "Should we cut pricing by 20% by Q3?" not "Should we think about pricing?"
Binary or small-N choice — "YC $500K at 7% vs $3M angel at $15M cap" not "What's the best funding option?"
Irreversible or hard-to-reverse — deprecating a product, relocating, hiring a VP, taking a term sheet
Stakes stated — "Should a 50-person startup hire 4 salespeople?" (8% of company, real commitment) not "Should we hire a salesperson?"
Second-order effects matter — if the answer depends only on first-order math, use a spreadsheet

Bad shape

Factual questions — "What's the SaaS discount industry standard?" → single LLM call
Pure forecasting — "What will the market do in 2027?" → autodecision chooses between actions, it doesn't predict the world
Low-stakes daily choices — the 15-minute loop overhead isn't worth it
Decisions you already have high conviction on — this will just slow you down and won't change your mind

Full loop examples

5 personas, 2 iterations, adversarial stress-testing, ~15 minutes.

Law firm AI replacement

"Should a mid-sized law firm replace all first-year associates with Claude + senior review?"

Full brief → · Deck (PDF) → · PPTX

What the council surfaced:

Full replacement was unanimously ranked last by all 5 personas — pipeline rupture in year 3, malpractice risk, and ranking damage all compound
The winning recommendation didn't exist in any single persona's output — it was synthesized from three independent alternatives (pessimist + customer advocate + regulator)
Regulator surfaced Utah vs Arizona sandbox distinction, WARN Act compliance, and MRPC 5.3 supervision analysis
Adversary found a correlated risk: the year-3 scenario where partners must become rainmakers IS the scenario where Claude has improved 30% more, collapsing the economic case for partial replacement too

Final recommendation: wrapped 18-month pilot cutting the 1Y class 50-60% with binding kill criteria, anchor-client co-design, and regulatory sandbox cover.

Buy vs rent vs relocate

"Dual-tech-income couple evaluates buying a house in Bangalore (10 Cr), renting + investing the delta, or relocating to San Francisco."

Full brief → · Deck (PDF) → · PPTX

What the council surfaced:

Council generated a creative alternative ("buy small + invest the rest") that outperformed all three original options on stress resilience — this option did not exist in the original question
SF savings corrected from $150-250K to $80-120K after modeling actual CA taxes, childcare, and living costs. The naive framing was off by 2x.
Adversary exposed a correlated risk: the scenario where you lose H1B sponsorship IS the scenario where your Indian portfolio is also down 30%. These tail risks are not independent.
Fragile insight identified: the buy-small option dominates only if liquidity stays above 6 months runway. Drop below and the option flips.

Final recommendation: staged real-options approach with buy-small as the default endpoint.

Adobe agentic CC vs UI-first Photoshop

"Should Adobe go all-in on agentic Creative Cloud and deprecate Photoshop's UI-first model within 3 years?"

Full brief → · Deck (PDF) → · PPTX

What the council surfaced:

The brief ran 3 days after Adobe publicly answered the question — Firefly AI Assistant launched April 15, 2026 as an additive layer, not a replacement. The data foundation flags this as CONFLICT-2, making the brief a stress-test of the road not taken.
Agentic inference COGS is the silent killer — agentic sessions cost 10-50x more compute than direct manipulation; at 41M subscribers even modest adoption creates substantial new COGS that none of the personas modeled at first
Regulator persona ranked #1 by peers — generated the two structural alternatives (dual-SKU + binding reversal triggers) that became the recommendation. Constraint analysts rarely drive creative synthesis; here they did.
The 3-year deprecation guarantees a "partial-agentic awkward state" at P=0.55 — agentic parity with 30 years of Photoshop workflow primitives is more realistic at 60-72 months. Once announced, deprecation is a one-way door (P=0.75 irreversibility).

Final recommendation: dual-SKU (Photoshop Pro with binding 10-year UI LTS + Photoshop Agentic), binding reversal triggers tied to NPS and Digital Media growth, quarterly user council of 50 working pros.

$50B buyback vs AI moonshot

"Should a $1T-cap big tech company spend $50B on share buybacks vs build a frontier AI moonshot?"

Full brief → · Deck (PDF) → · PPTX

What the council surfaced:

Pure $50B moonshot is eliminated by Reality Labs, not theory — Meta burned $69-77B cumulative on a comparable bet and still has under $5B revenue. The pattern, not the math, is the kill signal.
$50B is less than half of a single hyperscaler's annual AI capex ($120-200B) — a "moonshot" budget that's structurally underfunded against the actual frontier. Most readers assume $50B is big AI money. It isn't.
Five personas converged on the same hybrid synthesis from different starting points — Optimist via spin-out/JV, Pessimist via scaled buyback + partnership, Competitor via partnership-first, etc. The convergence across value systems is the strong signal.
The promoted path mirrors Microsoft-OpenAI specifically — $13B → $150B+ value creation. Survives 12/12 adversarial scenarios. Single binary gate: whether a frontier lab will accept minority equity without exclusivity (P=0.50).

Final recommendation: $10-15B equity in a frontier lab + $30-35B continued buyback + $5B vertical AI bet, with explicit kill criteria per tranche (no >$10B burn without $1B ARR, etc.) and a fallback to a higher buyback if the partnership gate doesn't open.

University optional in-person attendance

"Should a research university make in-person attendance optional across all undergraduate programs in the AI tutoring era?"

Full brief → · Deck (PDF) → · PPTX

What the council surfaced:

Universal optional attendance is regressive within a tuition tier — it redistributes value from first-gen and struggling students toward already-high-performers without lowering price. Counterintuitive and underexplored in public debate.
Regulatory cliff is the most decisive in the corpus — F-1 SEVP visa violation at P=0.85 risking $30-150M international revenue, and Title IV RSI reclassification at P=0.65 risking $50-300M federal aid. Same root cause: insufficient in-person structure. One structural fix defuses both.
De facto already optional, but the policy signal still matters — 73-83% post-COVID absence rates show students have voted, but formalizing what already happens still triggers regulatory and reputational consequences elite peers will weaponize as an admissions advantage (P=0.85)
Harvard 2025 RCT is the factual anchor — AI tutoring effect size 0.73-1.3 SD, the strongest learning-outcome data in the corpus. Cuts both ways: it justifies the question and shows what's at risk if implementation is botched.

Final recommendation: phased opt-in pilot (CS + Philosophy first, ~150 students), with quantitative kill criteria including >5% drop in first-gen retention. Six named action steps, eight monitor signals, seven explicit review triggers.

Comparison mode

10% vs 30% price cut

"Should a B2B SaaS drop pricing by 10%? By 30%? Which is worse?"

Full comparison → · 10% brief · 30% brief

Decks (8-slide quick mode): 10% deck (PDF) → (PPTX) · 30% deck (PDF) → (PPTX)

What the comparison surfaced:

Both reach "don't cut" but for fundamentally different reasons
10% fails by being too small — lands in the "dead zone" where enterprise procurement doesn't care but margins still compress 3-7 points
30% fails by being too large — needs 43% more volume to break even (near-impossible in B2B cycles), LTV drops 30%+ with compounding destruction, price anchor permanently resets
Both converge on the same alternative: restructure to hybrid/usage-based pricing (research shows 38% revenue growth premium)

The value of running both: identifying that "the right cut amount" is a false question. The underlying move should be restructure, not cut.

Decision types autodecision is good at

Decision Type	Why it fits
Pricing changes	Second-order effects dominate — competitor response, price anchoring, LTV, renegotiation
Market expansion	Multiple uncertain variables interact; irreversible once capital is deployed
Build vs buy	Asymmetric consequences over multi-year horizons; switching costs matter
Senior hiring	High cost to reverse, cultural second-order effects, opportunity cost of the slot
M&A / acquisitions	Black swan outcomes matter more than median; integration risk compounds
Fundraising terms	Dilution compounds; unusual terms have precedent effects on next round
Pivot or kill a product line	Sunk cost + team morale + external signal all interact
Relocation / career moves	Correlated risks (visa + market + family) are easy to miss individually
Strategic platform bets	Multi-year horizon, winner-take-most dynamics, deprecation risk

Decision types to avoid

Simple factual questions — a single LLM call is faster and cheaper
Low-stakes daily choices — the overhead isn't worth it
Decisions with already-high conviction — this will slow you down without changing your answer
Pure forecasting questions — autodecision picks between actions, it doesn't predict external outcomes
Questions with no clear action — "Should we think about X?" — scope it first, then come back

Try these

Copy-paste into Claude Code:

/autodecision "Should Adobe go all-in on agentic Creative Cloud and deprecate Photoshop's UI-first model within 3 years?"

/autodecision "Should Uber build their own autonomous vehicles instead of partnering with Waymo/Cruise?"

/autodecision "Should Netflix launch a free ad-supported tier in India, Brazil, and Indonesia?"

/autodecision:quick "Should we cut pricing by 20%?"
/autodecision "Should we cut pricing by 20%?"     # compare quick vs full on the same question

/autodecision "YC \$500K at 7% vs \$3M angel round at \$15M cap — which should a pre-revenue AI startup take?"

/autodecision "Should a 50-person startup hire 4 salespeople or spend the equivalent budget going AI-native on sales?"

/autodecision --template expansion "Should we launch in Southeast Asia?"
/autodecision --template build-vs-buy "Should we build our own auth system?"
/autodecision --template hiring "Should we hire a VP of Engineering?"

The loop (refresher)

1. Decompose decision      → sub-questions + constraints
2. Generate hypotheses     → competing paths, not a single narrative
3. Simulate outcomes       → 5 independent personas, first- and second-order effects
4. Critique assumptions    → anonymized peer review + adversary + sensitivity
5. Re-run with updates     → iterate until insights stabilize
6. Converge                → stable insights vs fragile insights + decision boundaries

See the README for the full 10-phase breakdown, persona details, and output format.