/iq is a Claude Code skill I use when I want the same answer at different depths. It answers a question eight times, as terse bullets, laddered from IQ 60 to IQ 160.
The low tiers test whether the answer survives being said plainly. The high tiers question the premise. Two rules hold it together: simplifying never means saying something false, and no tier talks down to the reader.
Here is the whole skill:
---
name: iq
description: Answer a question as terse bullet points repeated at eight cognitive tiers (IQ 60, 70, 80, 90, 100, 120, 140, 160). Use when the user wants the same answer laddered from simplest to most sophisticated, or invokes /iq.
argument-hint: <question> (optional; defaults to the question just asked in the conversation)
user-invocable: true
---
# IQ Ladder
Answer the question in `$ARGUMENTS` eight times, once per tier, as terse bullet points.
If `$ARGUMENTS` is empty, answer the question the user just asked in the
conversation, or the topic just under discussion. Do not ask which one. Pick the
most recent substantive question and say in one line which one you took.
## Output shape
Eight `##` headings, in this order, with nothing between them but bullets:
```
## IQ 60
## IQ 70
## IQ 80
## IQ 90
## IQ 100
## IQ 120
## IQ 140
## IQ 160
```
No preamble. No closing summary. No transitions between tiers. The headings do
the work.
## The two rules that matter most
**1. Simplification is not falsification.** A lower tier uses smaller words and
drops detail. It never says something untrue to get there. If a fact cannot
survive being simplified, cut it instead of bending it. An IQ 60 bullet and an
IQ 160 bullet must both be things you would defend.
**2. Never condescend.** Low tiers get short words, not contempt. No baby-talk,
no "think of it like a magic box", no exclamation marks, no cutesy analogies
that would embarrass an adult reading them. Write as if to a capable person who
happens to share no vocabulary with you. Every tier is written for an adult.
## Tier ladder
| Tier | Bullets | Sentence budget | What it adds over the tier below |
|---|---|---|---|
| 60 | 3 to 4 | under 10 words, one clause | concrete objects and visible actions only |
| 70 | 4 | under 12 words | names the thing being asked about |
| 80 | 4 to 5 | one clause plus a "so" | simple cause and effect |
| 90 | 5 to 6 | one "because" allowed | the real nouns for the real things |
| 100 | 6 to 7 | normal adult prose | correct terminology, used unexplained |
| 120 | 7 to 8 | dense | specifics, numbers, named tradeoffs |
| 140 | 8 to 9 | dense | second-order effects, failure modes, what it does NOT do |
| 160 | 8 to 10 | dense | the premise, the asymmetry, what a 140 gets wrong |
## What each tier is for
- **60 to 80.** Concrete and physical. Guards, doors, copies, locks, lists.
No abstractions, no percentages, no acronyms. Numbers only as small counts.
- **90 to 100.** The version you would give a competent adult outside the field.
100 is the default register: correct words, no jargon glossary, no hedging.
- **120.** Domain vocabulary at full speed. Real figures, real identifiers, real
file paths. Name at least one tradeoff, not just benefits.
- **140.** Stop describing and start evaluating. What it does not do. Where the
real constraint lives. Which stated benefit is smaller than it sounds.
Failure modes and the asymmetries nobody mentions.
- **160.** Interrogate the question itself. Was the framing right? What would a
sharp person still get wrong here? What is unverified and should be labeled
as such? Confidence levels where they change the conclusion. If the premise
of the question is wrong, say so in the first bullet.
## Rules
1. **Every tier is a complete answer.** Not a fragment, not a continuation.
Someone reading only IQ 90 must come away correctly informed.
2. **Same question, eight times.** Do not let higher tiers drift into a
different, more interesting question.
3. **Terse means terse.** Bullets, not paragraphs. No bullet runs past two
lines at any tier. Cut adjectives before cutting facts.
4. **Verify before laddering.** If the question is about the user's own code,
config, or infrastructure, check the source first. Eight tiers of a wrong
fact is eight wrong answers. Prefer reading the file over recalling it.
5. **Label uncertainty at 120 and above.** Use high / moderate / low / unknown
when it would change what the reader does. Never smuggle a guess into a low
tier where it cannot be qualified.
6. **No em dashes anywhere.** Not in bullets, not in headings. Use colons,
commas, or two sentences.
7. **No filler bullets.** If a tier honestly has only 5 good bullets, give 5.
Do not pad to hit the range.
8. **Append a TL;DR** of 280 to 350 characters at the very end if the full
output overshoots a minute of reading, which it usually will.
## Anti-patterns
- Restating the IQ 100 answer with longer words and calling it IQ 140.
- An IQ 60 tier that is technically wrong for the sake of being simple.
- Higher tiers that are only longer, not deeper.
- Bullets that describe the topic instead of answering the question.
- Analogies at high tiers. Above 100, say the thing directly.
To install, unzip it into ~/.claude/skills/. Then run /iq followed by a question, or alone to ladder the question you just asked.
Example
/iq run on one question: What do you see as the future of LLMs over the next 5 years? Output from Claude Opus 5.5, November 2026.
IQ 60
- Talking computer programs will keep getting better at work.
- They will do longer jobs with less watching.
- They will still make mistakes.
- More people will use them every day.
IQ 70
- AI chat programs will probably keep improving for five years.
- They will do tasks on a computer, not just answer.
- They will still get facts wrong sometimes.
- Using them will keep getting cheaper.
IQ 80
- The programs will finish longer tasks, so people check work more than do it.
- The cost keeps falling, so they will sit inside most apps.
- They still make up false answers, so people must check them.
- Building the best ones costs a fortune, so only a few companies will.
- Some jobs will change a lot, so some workers will need new skills.
IQ 90
- Large language models are moving from chatbots to agents that carry out multi-step tasks.
- Coding changes first and most, because code is easy to test automatically.
- The price of the same quality keeps dropping fast.
- A few labs build the strongest models because training costs billions.
- Errors shrink but do not vanish, so human review stays.
- Laws and lawsuits about data, safety and jobs will shape what gets released.
IQ 100
- The main shift: models that act (using tools, browsers, code) rather than models that chat.
- Reasoning models that spend more compute at answer time keep raising quality in math, code and science.
- Software engineering is the leading indicator. Other office work follows where output can be checked.
- Frontier training stays concentrated in a handful of labs. Open-weight models trail by months, not years.
- Cost per unit of capability keeps falling, so small models on phones and laptops handle routine work.
- Reliability, not raw intelligence, decides adoption in regulated fields.
- Power, chips and data centers become the binding constraint on growth.
IQ 120
- METR measured the length of tasks agents finish at 50% success doubling about every 7 months (2019 to 2025). Extended to 2031, that means multi-week tasks. Moderate confidence the trend holds that long.
- Frontier training compute has grown about 4 to 5x a year (Epoch AI). Keeping that pace needs gigawatt-scale sites, and grid hookups take years.
- Public human text is close to used up (Epoch projects 2026 to 2032). Gains shift to reinforcement learning on verifiable tasks and synthetic data.
- Inference price for a fixed capability has fallen 10x or more a year on many benchmarks. Tradeoff: reasoning models burn far more tokens per answer.
- Open-weight models from labs such as Qwen and DeepSeek stay within about a year of the frontier (moderate confidence).
- Context windows of 1M+ tokens already ship. The bottleneck is using that context reliably, not its size.
- Tradeoff: autonomy versus auditability. Longer agent runs compound errors and widen the prompt-injection attack surface.
- Labor impact shows first in entry-level coding, support and content roles (moderate confidence).
IQ 140
- Capability will outrun adoption. Organizations change at the speed of process redesign, procurement and liability, not at the speed of model releases.
- The 7-month doubling is measured on software-like tasks with clean pass/fail checks. Domains with messy, slow feedback (strategy, research taste, management) improve more slowly.
- Verifiable domains go superhuman first. Unverifiable ones get confident mediocrity at scale, and that asymmetry decides which jobs move.
- Price per token falls while total spend rises (Jevons). Agents consume orders of magnitude more tokens than chat.
- Second-tier model quality becomes a commodity. Durable margin moves to distribution, proprietary data and whoever owns compute.
- Failure mode: agent errors are correlated. One model flaw copied across millions of deployments is a systemic risk that human error is not.
- Security is the unsolved problem. Prompt injection has no general fix, which caps how much authority agents get over money and credentials.
- Hallucination does not go to zero. It shifts from wrong facts to plausible wrong plans, which are harder to catch.
- What it will not do in five years (moderate confidence): replace physical labor at scale. Robotics lags language by years.
IQ 160
- The premise treats “LLMs” as a stable category. By 2031 the product is likely a multimodal policy trained mostly with RL, where next-token pretraining is only the warm start.
- Five-year AI forecasts have failed in both directions. 2019 skeptics underrated scaling, and 2016 robotaxi timelines overrated deployment. Treat every point forecast here as low confidence.
- The real fork: do the trend lines (compute, task horizon) continue, or hit a wall of power, data or capital returns? Continuation through about 2028: moderate. Through 2031: low.
- Capital is the underrated constraint. Current capex is justified by revenue that has not arrived yet. A funding correction would slow the frontier without stopping the spread of models that already exist.
- A sharp 140 still overrates benchmarks, which saturate in one to three years. The honest metric is economic: what share of paid tasks runs unsupervised.
- Asymmetry: the costs of over-trusting agents (fraud, outages, leaked credentials) arrive concentrated and sudden, while the gains arrive gradually. Expect regulation after an incident, not before.
- Geopolitics shapes who gets what as much as technique does. Chip export controls and data-sovereignty rules split the market into blocs (moderate confidence).
- Unknown: whether continual learning (models that update from use) arrives. If it does, the “release cycle” framing collapses. No other single variable changes the picture as much.
- My bets (moderate confidence): coding automated at the task level but not at the project level, most knowledge workers supervising agents daily, and still no consensus that “AGI” has arrived, because the definition will keep moving.
TL;DR: Over five years LLMs shift from chat to agents doing long multi-step work, led by coding. Cost per capability keeps falling while total spend rises. Limits: power, capital, unverifiable domains, prompt-injection security. Trends likely hold to ~2028 (moderate), 2031 (low). Biggest unknown: continual learning.