autoresearch

Autoresearch (generalized)

Karpathy's loop in one breath: pick a goal, define a mechanical fitness function, constrain a mutable surface, iterate keep-or-revert until budget or plateau. This skill ports it to any domain.

The hard-won lesson: most candidate workflows are not autoresearch-shaped. Running the loop on a fat-tailed, reflexive, or slow-feedback problem converges to a local optimum of the wrong thing - very efficiently. This skill's main job is to catch that before the loop starts. Saying "this isn't autoresearch-shaped, here's why, here's what to do instead" is the correct output when a candidate fails triage.

When this skill applies

Invoke when the user:

• wants to scaffold or run an iterative improvement loop on any workflow
• mentions Karpathy, autoresearch, keep-or-revert, overnight optimization, Tobi's Liquid, pi-autoresearch, "the Karpathy loop"
• asks "can I apply autoresearch to X"
• has a measurable goal and wants autonomous iteration
• says "overnight agent," "experiment loop," "fitness function," "val_bpb"

Do NOT invoke for:

• one-shot optimization (no iteration needed - just do the thing)
• pure subjective quality with no path to a mechanical metric and no willingness to build one
• safety-critical / irreversible actions where keep-or-revert cannot be automated
• architecture or strategy questions (use the council skill instead - "use autoresearch for plumbing, never for architecture")

Simplicity principles (Karpathy, verbatim)

Baked into every stage. Non-negotiable. Quoted directly from Karpathy's program.md:

1. One atomic change per iteration. Not five. Not "a batch of related tweaks." One. Revert is trivial; diagnosis is tractable.

2. Small mutable surface. Karpathy's original had exactly ONE file mutable (train.py) and one file read-only (prepare.py). If your surface is "the whole codebase" or "the whole prompt," the loop cannot converge. Scope down before looping.

3. Simple fitness function. One command. One number. Fixed corpus. Karpathy's was grep "^val_bpb:" run.log. If your fitness needs a 200-line harness, your fitness is wrong.

4. Simple loop driver. ~80 lines of bash, or one prompt that "LOOPs FOREVER" in a single session. Not a framework. Not microservices. Karpathy's entire system is train.py + prepare.py + results.tsv + program.md - four files.

5. The simplicity criterion (Karpathy's exact rule for keep-or-revert decisions):

   "All else being equal, simpler is better. A small improvement that adds ugly complexity is not worth it. Conversely, removing something and getting equal or better results is a great outcome - that's a simplification win. When evaluating whether to keep a change, weigh the complexity cost against the improvement magnitude. A 0.001 val_bpb improvement that adds 20 lines of hacky code? Probably not worth it. A 0.001 val_bpb improvement from deleting code? Definitely keep. An improvement of ~0 but much simpler code? Keep."

   Decision table:

   Fitness delta	Complexity delta	Decision
   Small improvement	Added ugly code	Revert
   Small improvement	Deleted code	Keep (simplification win)
   ~0	Much simpler	Keep (deletion alone is valuable)
   Big improvement	Added code	Keep (paid for by the gain)
   Big improvement	Deleted code	Definitely keep

   This is NOT a rule about change magnitude ("be bold"). It is a rule about the trade-off between complexity cost and fitness benefit. Simplifications always win, even at zero fitness improvement.

6. When stuck, escalate ambition - not iteration count. Karpathy: "If you run out of ideas, think harder - read papers referenced in the code, re-read the in-scope files for new angles, try combining previous near-misses, try more radical architectural changes." If timid tweaks plateau, the answer is BOLDER changes, not more timid ones. See references/anti-patterns.md #13.

7. Hard iteration cap + autonomous execution. The loop driver enforces a max iteration count (default 50) and a dollar budget. Within those limits, the loop runs AUTONOMOUSLY - never pausing to ask permission. Karpathy, verbatim:

   "Once the experiment loop has begun, do NOT pause to ask the human if you should continue. Do NOT ask 'should I keep going?' or 'is this a good stopping point?'. The human might be asleep. You are autonomous. The loop runs until the human interrupts you, period."

   The cap is a termination condition, not an "ask the human" event. The loop stops when: cap hit, budget exhausted, plateau detected (10 iterations without significance-threshold improvement), guard violation, or manual interrupt. Nothing else.

These are why Karpathy's autoresearch.py is ~630 lines, not 6300. Complexity is the enemy of convergence.

Five stages, in order

1. Intake - what, how measured, what can change, what can't regress, budget
2. Triage - is this problem autoresearch-shaped? (see references/triage-checklist.md)
3. Fitness design - build or validate the fitness function (see references/fitness-design.md)
4. Mode selection - pure / barbell / via negativa / inverted / human-in-loop (see references/modes.md)
5. Loop + post-mortem - scaffold, run, diagnose plateau if reached (see references/anti-patterns.md when diagnosing)

Skip stages only when the user explicitly says "I've done that, start the loop."

Reference files map:

• references/triage-checklist.md - Stage 2 (is this problem autoresearch-shaped)
• references/fitness-design.md - Stage 3 (the 7 requirements a fitness function must meet)
• references/modes.md - Stage 4 (pick one: pure / barbell / via negativa / inverted / human-in-loop)
• references/anti-patterns.md - Stages 3 + 5 + 6 (14 named failure modes)
• references/plateau-ideation.md - Stage 5/6 (what to propose when the loop plateaus)

Load each reference file on demand during the stage that needs it. Do not preload.

---

Stage 1: Intake

Ask all five in one structured message, not one at a time:

1. Goal - what should be better after this runs? State as "before → after."
2. Candidate metric - how would you know it's better? A number is ideal; a deterministic rubric is the fallback.
3. Mutable surface - what files, prompts, config, copy, or parameters can the agent change? Be specific.
4. Guard - what must not regress? Safety tests, brand rules, legal constraints, downstream dependencies.
5. Budget - iterations, wall-clock, or dollars.

If the user cannot answer 2 or 3, continue anyway. Triage will surface it and Stage 3 will force a resolution.

---

Stage 2: Triage

Load references/triage-checklist.md. Walk the candidate through five dimensions and score each green / yellow / red:

• Feedback latency
• Metric mechanicality
• Tail shape (thin vs fat-tailed)
• Sample size vs noise
• Surface locality (atomic diff vs system redesign)

Decision rules:

• All green → pure optimizer mode, proceed to Stage 3
• Any yellow → mode adaptation required, proceed to Stage 3 with concerns noted
• Any red → either adapt the mode or refuse the loop and offer alternatives (council, human-in-loop, fix the upstream bottleneck first, smaller better-shaped sub-problem)

Refusal is a feature. Do not start the loop to be agreeable.

---

Stage 3: Fitness design

Load references/fitness-design.md. The fitness function is where autoresearch dies when it dies - not in the loop mechanics.

Required before the loop can start:

1. Fitness is a command (or deterministic chain) that outputs a number
2. It runs on a fixed corpus with fixed seeds (stationary across iterations)
3. Baseline has been recorded to baseline.json
4. Significance threshold defined (what delta is real vs noise)
5. Sanity check done: score known-good and known-bad by hand, confirm the metric agrees
6. Guard metrics defined: what must not regress

If 1-4 are not met, the skill's answer is: "fix the fitness function first, then come back." This is the single highest-leverage intervention. Do not skip it to be helpful.

---

Stage 4: Mode selection

Load references/modes.md. Pick exactly one:

Mode	When
Pure optimizer	Small mutable surface, fast feedback, clean metric, thin-tailed
Barbell	Fat-tailed, reflexive domain - 85% immutable proven baseline + 15% wild experiments
Via negativa	Fragile system - ask "what to kill" not "what to add"
Inverted	Imbalanced classes - optimize rejection of bad, not selection of good
Human-in-loop	Slow feedback or irreversible actions - agent proposes, human decides

Record the mode in goal.md. It controls what the loop is allowed to propose.

---

Stage 5: Loop + post-mortem

Scaffold (once)

Create in the user's working directory (or a subdirectory they pick):

• goal.md - filled from templates/goal.md: goal / metric / surface / guard / mode / budget
• The fitness command (script, notebook, evaluator - user's choice, not the skill's)
• experiments/ directory for per-iteration logs
• baseline.json with the starting metric value and sanity check results
• loop-driver.sh and iteration-prompt.md from templates/ (for autonomous runs)

Drive the loop (autonomously)

Claude Code's interactive mode pauses between turns. To run the loop continuously without a human pressing enter each iteration, drive it from OUTSIDE the Claude Code session via a script. Each iteration is one fresh claude -p (non-interactive print mode) invocation that reads state from goal.md + experiments/, proposes one atomic change, runs fitness, commits or reverts, appends to journal.md, and exits.

Three options, in order of simplicity:

1. Bash + claude -p (recommended, simplest). Use templates/loop-driver.sh and templates/iteration-prompt.md. Run:

   ./loop-driver.sh --max-iter 50 --max-budget-usd 10
   

   Uses Claude Code's -p / --print mode, --permission-mode acceptEdits for tool autonomy, --max-budget-usd for a built-in dollar cap. State lives in git + files, not session memory. Each iteration is a fresh session reading the journal.

2. Python + Anthropic SDK - what Karpathy's original autoresearch.py does. Use when you need programmatic state tracking, richer plateau detection, or parallel experiments. More code, more control.

3. Interactive / manual - user approves each iteration in a single Claude Code session. OK for N ≤ 5. Tedious above that.

Hard caps are mandatory. Every driver MUST enforce:

• Max iterations - default 50, hard cap. Raise only with a written reason in goal.md.
• Dollar budget - via claude -p --max-budget-usd.
• Wall-clock budget - via timeout or driver-level check.
• Plateau detection - default 10 iterations without significance-threshold improvement → stop.
• Guard violation - immediate stop, no retries.

Any trigger writes experiments/STOP with a one-line reason and exits. Stage 6 post-mortem reads this file and produces the plateau diagnosis.

What does NOT work: Claude Code hooks (Stop, PostToolUse, UserPromptSubmit, SubagentStop). Hooks REACT to Claude's actions - they cannot inject new prompts that drive iteration. The loop driver MUST live outside the Claude Code session.

Run the loop

1. Read state: git log, experiments/, last N lessons learned
2. Propose ONE focused change, respecting the mode
3. Commit the proposal with an experiment: prefix BEFORE running (so git can revert)
4. Run the fitness command
5. Compare to baseline (or best-so-far) with the significance threshold
6. Keep (advance baseline) or revert (git revert)
7. Append a one-line lesson to experiments/journal.md
8. Check termination: budget exhausted, plateau detected (10 iterations without improvement), guard violated → Stage 6

Non-negotiable principles:

• One change per iteration
• Mechanical verification only (no "looks good")
• Automatic rollback on regression or guard failure
• Git is memory

Post-mortem (when the loop stops)

Produce:

• Final metric vs baseline, with significance
• Top 3 changes that stuck, top 3 changes that were reverted
• Plateau diagnosis if plateaued - consult references/anti-patterns.md and name one:
  • Wrong search space → change the mutable surface
  • Wrong metric → fitness does not correlate with the real goal
  • Broken measurement → sample size, stationarity, or confounded labels
  • Genuine local optimum → switch mode to barbell
  • Out of budget → more iterations would help

Before declaring plateau, run the ideation workflow in references/plateau-ideation.md. This is mandatory, not optional. Specifically:

1. Mine the last 10 reverts for pattern (axis + failure mode for each). Tradeoff reverts (primary lifted, guard broke) are especially rich - they reveal axes with slack that need compensating moves elsewhere.
2. Run the taxonomy coverage check. If fewer than ~6 of 8 axes have been touched with substantive changes, the loop is NOT plateaued - it has exhausted one corner of the search space. Propose on an untouched axis and continue.
3. Only when coverage is satisfied AND revert mining surfaces no untouched direction: invoke /council with the three structured prompts from plateau-ideation.md Move 3.
4. Only when all three are exhausted: declare plateau and pick a cause above.

"I ran out of ideas" is anti-pattern #13 (timid tweaks). The plateau-ideation workflow is how the skill converts that into a concrete next move instead of a stop.

The plateau diagnosis IS the deliverable when the loop doesn't improve. Foreground it, don't bury it.

Escalation: invoke /council on a plateaued loop

When the loop plateaus and you own the mutable surface, the mode, and the fitness function, the failure is usually invisible from inside the loop - wrong sample size, unverified compliance, a metric that doesn't discriminate, confounded labels, or a mutable surface that no longer holds leverage. These are exactly the findings a multi-perspective review delivers cheaply in one pass.

Before switching modes or mutable surfaces, spend a single /council call. Canonical prompt:

Goal, fitness, baseline, significance threshold, mode, last 10 journal lines - the loop has plateaued. What is the loop missing that an outsider would notice?

A council that agrees with your reading tells you the loop is genuinely stuck on the right problem. A council that disagrees usually names the specific failure mode from references/anti-patterns.md - which is exactly the plateau diagnosis you were about to struggle to produce alone.

This is different from anti-pattern #10 ("use council for architecture, not tactics"). Council here is a second-opinion tool on the loop itself, not an architecture substitute. On a real autoresearch run, a plateau-council review produced the two highest-leverage fixes (enlarge sample size, verify compliance) that the loop couldn't see from inside.

---

Style rules for this skill

• Declarative: tell the user WHAT to check, not exact commands. The agent picks the command.
• Refuse gracefully and suggest alternatives when a candidate fails triage
• Keep the loop boring - creativity lives in triage and mode selection, not loop mechanics
• Git is the experiment log; journal.md is the narrative
• Use plain - hyphens, not em dashes