You're shipping faster than ever. Are you getting sharper, or just getting carried?
The crux is the part of the work that's still yours: what you reject, redirect, and refuse to ship. Crux captures it, so you can see whether your judgment is sharpening or quietly eroding while the AI does the typing.
This is ongoing research, not a product. The instrument in this repo is the study's method and your replication path; the document is the work.
Start here: the research document — a living field study of the human side of AI collaboration, and what I found measuring my own work. v2 extended it from self-knowledge to work itself: the ownership account from the field, the labor question (people keep the job and lose the capability, and no hiring signal can see it), and a memo to the platform layer. v3: the same week that memo went live, Anthropic shipped Reflect — the first platform instrument for the human side. It measures usage. The judgment half stays open, and this research keeps measuring it.
You commit code. You close tickets. You ship features. Your output has never been higher.
But output isn't the same as judgment. And right now, nothing measures whether the human behind the AI is improving, plateauing, or quietly atrophying.
Anthropic studied nearly 10,000 conversations and found that when AI produces polished output, users become less critical — they stop questioning, stop fact-checking, stop pushing back. They call it the artifact paradox. The better the AI gets, the easier it is to coast.
When they interviewed 80,508 Claude users, cognitive atrophy was the #4 concern. 13,000 people explicitly worried about it. Nearly half already felt it.
"I don't think as much as I used to. I struggle to put the ideas I do have into words." — Heavy AI user, United States
Wharton has now codified the umbrella term: cognitive surrender. Ethan Mollick's Choosing to Stay Human (May 2026) makes the case that the answer isn't user-level willpower but system-level constraints. This is the gap crux exists to surface.
This isn't about using AI less. It's about knowing whether you're still thinking while you use it.
Anthropic's 4D framework — Delegation, Description, Discernment, Diligence — defines baseline AI competency. Courses will teach it. Everyone will pass. It stops being a differentiator fast.
What separates people who get better with AI from people who get carried by it is a layer above fluency:
- Trust calibration — knowing when to let the AI run vs when to check its work. Task-by-task, not blanket.
- Artifact paradox resistance — the more polished the output looks, the harder you examine it.
- Meta-cognition — correcting the AI's process, not just its output.
- Strategic killing — deciding what not to build, even when the AI builds it well.
- Pattern recognition — seeing your own habits across sessions, not just within them.
These are dispositions, not skills. You develop them through practice — but only if you can see them. Your GitHub shows what got committed. It doesn't show what you rejected, redirected, or killed.
A practice for developing AI judgment. Like a training log for a runner — the primary value is to you.
Zero commands. Zero friction.
A Claude Code hook fires when every session ends. It reads the transcript, extracts collaboration patterns, and generates a digest automatically. You don't invoke anything. You don't remember anything. You just work.
For manual sessions, /marmite still works — type it at the start, say "wrap" when you're done, and Claude silently tracks the moments where you redirect, reframe, set a quality bar, or kill a direction.
# 2026-02-23 — eval scoring diagnosis
~4h research session. Suspected a generation problem, investigated before fixing.
- Instinct to diagnose before treating — scores looked bad but the question
was whether the answers were wrong or the measurement was wrong. Chose to
investigate the scorer first.
- Conservative fix philosophy: built targeted matchers with guards that would
rather miss a match than fabricate one. Precision over recall.
- Refused to chase a low score when the test conditions were unreliable.
Flagged it as bad data instead of reacting to the number.
- Held the fix as research rather than shipping immediately. Won't merge
until re-validated under correct conditions.
- Pattern: diagnose root cause before touching anything, don't ship under
uncertainty, resist the urge to fix what isn't broken.Skimmable in 15 seconds. No scores. No self-assessment. Just a record of how someone actually works through a problem with AI.
Individual digests capture sessions. The real signal is in the trajectory across them.
Run /trail periodically to generate a synthesis:
# Synthesis — February 2026 (3 sessions)
## Recurring patterns
- Consistently diagnoses root cause before fixing. Two sessions where the
obvious fix was wrong and investigation saved significant rework.
- Delegates mechanical execution freely but always verifies judgement calls.
Trust is calibrated by task type, not blanket.
## Evolution
- First session: accepted AI-generated digests at face value. By third
session: caught the digest leaking architecture and rewrote the skill.
Meta-awareness of AI output quality is increasing.
## Beyond fluency
- Artifact paradox resistance: actively challenges polished output in 2/3
sessions. This is the behaviour Anthropic's research shows most users
stop doing.The synthesis surfaces whether your collaboration instincts are sharpening, plateauing, or eroding. You can't course-correct a disposition you can't see.
- Engineers already building with AI who want a feedback loop for judgment, not just code.
- Tech leads who need to develop their team's AI collaboration instincts, not just their prompt skills.
- Teams where tacit knowledge is disappearing — digests are a knowledge-transfer mechanism (juniors can see how seniors work through a problem) and a retro substrate: a way to retro the collaboration, not just the product.
The primary reader is always the person who runs it. It's a practice log, not a scoreboard and not a hiring signal — there are no scores and no cross-user comparison by design.
crux today is a Claude Code-only practice tool. The position it occupies is the human-side reflection layer for AI collaboration — a parallel to what providers are building on the AI side (Anthropic's Auto Dream, persistent memory; OpenAI's chat memory; etc).
The endgame is not a SaaS. The research document closes with a memo to the platform layer: the instrument shipped native to the tool, a disclosure moment at sign-off of autonomous runs, and an opt-in longitudinal panel — the individual-level counterpart to population usage indices like the Economic Index. This repo is the reference implementation and the head start. No patents, on purpose.
The roadmap (see ROADMAP.md) covers the depth work (the Trail, delegation sign-off), the team layer (workflow ownership, the Retro), and cross-LLM capture (Cursor, Claude.ai exports, Copilot OTel, ChatGPT exports).
The strongest response to this research is a replication. Install takes a few minutes; after 30 days you'll have your own shape to compare against mine. If you publish yours, tell me — the document cites replications.
Clone the repo and hook it into Claude Code:
git clone https://github.com/ElliotJLT/crux.git ~/crux
chmod +x ~/crux/scripts/auto-digest.sh ~/crux/scripts/digest-worker.shThen add the SessionEnd hook to your global Claude Code settings (~/.claude/settings.json):
{
"hooks": {
"SessionEnd": [
{
"matcher": "",
"hooks": [
{
"type": "command",
"command": "~/crux/scripts/auto-digest.sh"
}
]
}
]
}
}That's it. Every session with 4+ user messages now generates a digest automatically when you exit. Short sessions and trivial ones (config, installs, quick lookups) are skipped. Digests land in decisions/digests/.
Each digest is automatically classified using Anthropic's Economic Index framework — tagging the interaction type (directive, feedback-loop, task-iteration, validation, learning) and collaboration mode (automation vs augmentation).
Requires: jq, python3, and claude CLI in your PATH. The install script checks for these.
Custom install path: Set CRUX_DIR to override the default ~/crux location.
Opt out of auto-capture for a specific project by dropping an empty file in its root:
touch /path/to/project/.crux-skipOr skip globally for a session with CRUX_SKIP=1.
mkdir -p ~/.claude/commands
curl -o ~/.claude/commands/marmite.md \
https://raw.githubusercontent.com/ElliotJLT/crux/main/.claude/commands/marmite.md
curl -o ~/.claude/commands/trail.md \
https://raw.githubusercontent.com/ElliotJLT/crux/main/.claude/commands/trail.md/marmite gives richer in-session tracking (it observes as you work, not post-hoc from the transcript). /trail generates monthly synthesis. Both work in every project.
Once you have 5+ digests, generate a trajectory synthesis:
~/crux/scripts/synthesize.shThis produces a synthesis in decisions/synthesis/YYYY-MM.md that shows:
- Your augmentation/automation split vs Anthropic's 59% augmentation benchmark
- Which interaction types you use most (and which you're missing)
- Trend direction — improving, flat, or declining
- Coasting alerts if you're drifting toward directive-heavy sessions
Run it weekly via cron, or just run /trail which covers the same ground.
Installed from the cloned repo (not yet published to PyPI):
cd ~/crux && pip install .
# Your collaboration patterns and trend (descriptive, no score)
crux metrics
# Render your Collaboration Log (markdown + HTML)
crux profile --format both
# Preview locally
crux serveEveryone will be fluent. Anthropic can measure it, courses can teach it, and the floor is rising fast enough that it stops being a differentiator. But population-level measurement (how do people use AI on average?) is different from individual measurement (how do you use AI, and is it changing?). crux bridges that gap.
Anthropic is already teaching people to write AI diligence statements — transparent disclosures of what AI did, what you did, and how you verified the result. It takes 10 minutes per deliverable when you do it manually.
What can't be taught is the layer above fluency: the instinct to challenge, the discipline to kill, the judgment to know when the AI is confidently wrong. Ferrier calls it Learning Power. Shumer calls it the thing that makes you irreplaceable. Osmani calls it taste.
crux captures that layer automatically — by making it visible to the person who needs to see it most: you.
- Ethan Mollick: Choosing to Stay Human (May 2026) — Codifies Wharton's "cognitive surrender" as the umbrella term. Argues the answer is system-level constraints, not user-level willpower. The intellectual anchor for what this tool is for.
- Anthropic: Auto Dream — Managed Agent Memory Consolidation (May 2026) — Anthropic's own memory layer for AI. crux is the parallel for the human side.
- Anthropic: What 81,000 People Want from AI — Cognitive atrophy is the #4 concern. 13,000 users worried about it. Half already feel it.
- Anthropic: Writing an AI Diligence Statement — The manual version of what crux automates
- Anthropic: The AI Fluency Index — Why polished output makes people less critical
- Anthropic: The Anthropic Economic Index — How AI is used in the economy: automation vs augmentation, 5 interaction types, and the benchmark data crux classifies against
- Anthropic: The Impact of AI on Computer Science Education — How AI changes skill acquisition and the case for deliberate practice
- Victoria Ferrier: AI Is The Wrong Unit of Analysis — Human capacity, not AI capability, determines success
- Matt Shumer: Something Big Is Happening — The inflection point and why differentiation matters now
- Taste Is the New Bottleneck — Why judgment can no longer remain tacit
- Addy Osmani: The Next Two Years — What makes engineers valuable when AI codes
- Hiring Is Shifting From Syntax to Judgment — 54x increase in aptitude assessments
- Are We Automating Into a Knowledge Crisis? — The tacit knowledge gap
- Boris Cherny: What Happens After Coding Is Solved — Head of Claude Code on the builder role
- Software Engineers Face an Identity Crisis (Business Insider, via Deedy Das) — The craftsman tax: review burden lands on the most expert; the ownership account in the wild
- AI vs Gen Z: The Junior Developer Crisis — How stepping-stone tasks are disappearing
MIT