elon-algorithm

Elon Algorithm — cut AI slop, keep what's load-bearing

AI output is additive: it pads, hedges, over-explains, and invents requirements nobody asked for. Elon's algorithm is subtractive — the exact antidote. This skill runs it on any artifact and hands you a cut-plan to review.

The algorithm (the per-artifact filter)

1. Question every requirement — trace each part back to a real ask. If no human/prompt asked for it, it's a cut candidate. Requirements the model invented and stated confidently are the most dangerous.
2. Delete — default to removing; you add back only the ~10% that something concrete breaks without.
3. Simplify what survives.

(Steps 4–5, accelerate and automate, are about the factory, not the part — out of scope for a single-artifact pass.)

How the workflow works

It clones the artifact first (the original is never touched), then runs four phases:

• Bird's-eye pass — cross-file fate you can only see by looking across files: whole files/dirs that duplicate each other (mirror trees, vendored docs restating guidance), with each redundancy mapped to its canonical home.
• Per-file pass — one coarse outcome per surviving file: a leaner rewrite, a whole-file delete, or keep. (Files sealed for deletion above aren't re-simplified — don't polish a part you're removing.)
• Asymmetric debate — per proposed cut, a lone objection must name a concrete breakage (a lost instruction, the sole place a fact appears, a dangling reference); "it's useful / adds context" doesn't count. The deletion side gets the last word.
• Cut-by-default judge — one coherent knife. Keeps something only if a concrete objection survived the rebuttal. Returns a cut-plan + a ranked add-back menu (least-confident cuts first, for the human to restore their 10%).

The output is a plan, not a blind rewrite — you review it and apply it as a diff/PR, staying in the loop at the deletion step where the human belongs.

Run it

The four phases above are the spirit of the algorithm, not a tool. How you run them depends on the size of the artifact:

Small artifact (a single file, or a couple) — run it directly

When the target is one file or a small handful, you don't need the workflow — but the structural altitude still applies. The cross-file pass is moot with one file; questioning whether each section / block / requirement deserves to exist is not, and that's where the deepest cuts live. So attack structure before copy: default to deleting or merging whole sections, then simplify the survivors line-by-line. The failure mode is treating the existing structure as fixed and only trimming sentences inside it — a 1% pass, not an Elon pass. Then run the same clone → review → debate → judge shape with sub-agents directly (structural cuts first, then line-level; deletion gets the last word), no dynamic workflow.

This is the lighter, faster default for most "trim this one file/prompt/skill" asks.

Large or multi-file artifact (a directory, a bundle, a mirror tree) — use the workflow

When the target spans many files, invoke the bundled Workflow, which orchestrates the full fan-out deterministically:

Workflow({
  scriptPath: "${CLAUDE_PLUGIN_ROOT}/workflows/elon-algorithm.js",
  args: {
    path: "<file OR directory/bundle to simplify>",
    standardHint: "<the standard the artifact aspires to>"
  }
})

• path — a single file, or a whole directory/bundle (the bird's-eye pass handles multi-file artifacts). Use text instead for inline content with no file.
• standardHint (optional but high-leverage) — the taste/standard to judge against: a style rule, a skill's own doctrine, "every line must change agent behavior at a decision point", etc. The sharper the standard, the sharper the cuts.

It returns { cloneRoot, proposals, plan } — plan.decisions (per-target verdict + confidence) and plan.addBackMenu. Hand that to a fresh agent to apply as a PR; don't burn your own context materializing edits.

Adapt it

The workflow is a starting point, not a black box. Copy workflows/elon-algorithm.js, tweak the lenses, the phases, or the prompts for your artifact, and run your version. The reusable shape is clone → review → debate → judge; everything else is tuning.

When to use

• Trimming a bloated skill, prompt, system message, or doc.
• Reviewing AI-generated code or writing for slop before it ships.
• Any "make this leaner / denser / less over-engineered" ask on an artifact you can point at.