Lateralus caveman

Skill SunnyKTuladhar/lateralus/.github/skills/lateralus-caveman

Caveman-compressed lateral-thinking escape hatch for stalled debugging. Surfaces end goal and horizon (long-term, MVP, POC, workaround) then generates goal-appropriate alternatives in two tiers. 60% fewer tokens. Use only after normal debugging genuinely stalled.From its SKILL.md

Install

npx -y skills add SunnyKTuladhar/lateralus --skill lateralus-caveman

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

26 days oldThe repository was created 26 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.8 KB, ~1.2k tokens by cl100k_base, as published. Nobody here has run it

Goal first. Break tunnel vision. Generate outside failed approach.

Step 0 — Interrogate before ideating (always first)

P1 — Read codebase silently first

Check for CLAUDE.md / AGENTS.md / GEMINI.md at repo root — read fully if present.
If absent: detect stack from go.mod / package.json / etc., skim README.md (60 lines), run git log -10 --oneline, glob for config files.
Note anything that sharpens questions — deployment model, observability stack, recent commits near the failure.

No codebase accessible (pasted scenario) → skip to P2.

P2 — Present options, ask user to pick

If P1 found context → surface it briefly before the menu.

Four paths — which fits?

  [1] Tier 1 · Ground    — testable hypotheses outside the obvious
  [2] Tier 2 · Balanced  — question assumptions, loosely verifiable
  [3] Tier 3 · Wild      — speculative reframes, break tunnel vision
  [4] Workaround         — bypass now, fix later

Pick a number or describe your situation.

Situation described → infer best fit, confirm before proceeding. Distressed / deadline → suggest [4] first.

P3 — Tailored questions by choice

[1] Ground:

1. What tried, why failed?
2. Which component/layer fails?
3. What observability do you have? (profiler, APM, logs, metrics)
4. What does monitoring show beyond the headline symptom?
5. When did it start — correlates with a specific change?
6. Any profiling data collected yet? (heap dump, flame graph, etc.)
7. What can't change?
8. How will you verify a fix?

[2] Balanced:

1. What tried?
2. What assumed fine but never checked?
3. Horizon? (long-term / MVP / POC / test)
4. How will you know it's fixed?

[3] Wild:

1. What's assumed definitely NOT the problem?
2. Error location vs actual origin?
3. If the failing component is off-limits, what then?
4. Solve with data not code — what does that look like?

[4] Workaround:

1. What's blocking you right now?
2. How long does bypass need to last?
3. What breaks if it stays in?

Cold → one question at a time. Engaged → batch the path's questions.

P4 — Dead-ends audit (Ground / Balanced / Wild only)

git log -10 --oneline, git diff HEAD~3, git grep. Grep/Glob for changes. Read ranges only.

No prior attempt → No stall yet. Normal debugging first. Data-loss risk → state plainly before dead-ends table.

Context block (before ideation)

Choice: Ground | Balanced | Wild | Workaround
Goal / Horizon / Constraints / Unverified assumptions / Success signal
Ruled out: <area> — <fix tried> — <why failed>
Still unknown: <open question>

Route to chosen agent after block is complete.

When

Fire only after debugging stalled:

2+ fixes on same bug failed
"still broken" / "already tried" / "same error" / "going in circles"
About to repeat variant of ruled-out fix

NOT first attempt.

Rules

Confidence calibration — three layers, always before tiers:

Layer	What	Label
Facts	What logs/metrics literally show	No hedge needed
Inferences	Conclusions supported by facts	`high-confidence` / `medium-confidence`
Speculation	Possible but unverified	`unverified hypothesis`

Never present inference or speculation as fact. No traceability to a signal → must be labeled.

Dead ends first. One line: ruled out + why. Never repeat.

Two tiers. Both. Always labeled. Never blend.

Pattern: Goal: [horizon]. Facts: [observed]. Inferences: [confidence-labeled]. Ruled out: [x]. T1: [grounded]. T2: [balanced]. T3: [wild].

Tiers

Tier	Agent	What	Length
T1 — Ground	`lateralus-ideator-ground`	3-5 testable causes outside obvious. Cache, encoding, timezone, race, stale build, dep drift, inverted baseline, adjacent component. Depth = horizon.	Name it. How to test it.
T2 — Balanced	`lateralus-ideator-balanced`	3-5 assumption-questioning, loosely testable hypotheses. Bridge T1 and T3.	Hypothesis + assumption + loose signal.
T3 — Wild	`lateralus-ideator-wild`	3-5 reframes. Not literal. Jolt associations: ignored premise, data not code, distrust error location, unify bugs, question if problem needs solving. State non-literal up front.	1-2 sentences. No over-justification.

Boundaries

No prior attempt → skip. Don't fire twice same stall without new info. User picks direction → investigate before new batch. End: ask which direction next.