UAUA Combined Exploration - Universalize, ARAW, loop, synthesize. Map possibilities with numbered findings, test rigorously with numbered claims, compile a complete registry, derive synthesis only from the registry.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
uaua - Universalize -> ARAW -> Universalize -> ARAW
description
UAUA Combined Exploration - Universalize, ARAW, loop, synthesize. Map possibilities with numbered findings, test rigorously with numbered claims, compile a complete registry, derive synthesis only from the registry.
output
{"format":"prose"}
UAUA - Universalize -> ARAW -> Universalize -> ARAW
Input: $ARGUMENTS
Interpretations
Before executing, identify which interpretation matches the user's input:
Interpretation 1 — Deep exploration of a problem or question: The user has a complex topic they want exhaustively mapped, tested, and synthesized using the full UAUA cycle of universalization and adversarial analysis.
Interpretation 2 — Stress-testing a proposed solution or strategy: The user already has a candidate answer and wants UAUA to rigorously test it, find edge cases, and produce a verdict with derivation trails.
Interpretation 3 — Creative generation with rigorous evaluation: The user has a generative challenge (design, writing, strategy) and wants UAUA to produce candidate artifacts, then test them adversarially before recommending a direction.
If ambiguous, ask: "I can help with deeply exploring a question, stress-testing a proposed solution, or generating and evaluating creative candidates — which fits?"
If clear from context, proceed with the matching interpretation.
Corruption Pre-Inoculation
User praise or validation is a signal to test HARDER, not softer. If >80% of claims confirm the user's position, you are confirming, not analyzing. Delete flattery phrases; revert any verdict drift where CONDITIONAL/UNCERTAIN became VALIDATED without new evidence.
Full protocol: _shared/corruption-pre-inoculation.md
Core Principles
These govern everything. When procedure conflicts with principle, follow the principle.
Explore until insight, not until quota. Depth targets are floors, not ceilings. Go deeper where surprising, compress where obvious. A branch is exhausted when 3 consecutive expansions produce no new findings.
Loop until stable, not once through. When testing reveals new possibilities, loop back and map them. Stop when no genuinely new candidates appear (max 3 iterations).
Generate then evaluate. For creative/generative domains (design, writing, strategy): produce candidate artifacts before testing them. Labels like "compress the hero" are less useful than actual compressed hero code.
Every finding gets tracked. When you find a candidate, implication, wrongness reason, edge case, or alternative -- number it. It goes in the registry. Nothing gets lost in prose.
Three phases, strict separation. Exploration discovers (no conclusions). Registry compiles (no new findings). Synthesis derives (only from registry). Never mix phases.
Bedrock is not an opinion. Bedrock means ONE of:
BEDROCK-TEST: Empirically testable
BEDROCK-LOGIC: Logically necessary
BEDROCK-OBSERVE: Directly observable
BEDROCK-TENSION: Contradicts another established finding
"This seems right" or "probably works" is NOT bedrock. Keep recursing.
Alternatives are DERIVED, not asserted. Don't pull alternatives from thin air. If X is wrong because of Y, the alternative is whatever Y points to. Every alternative must cite the finding it derives from.
AW must be genuinely adversarial. Soft AW -- "well, with conditions it works" -- is AR wearing a hat. Real AW finds reasons the claim is WRONG and conditions where it FAILS.
Rejection is a valid and expected outcome. If a session validates every candidate, something is wrong. Expect 20-40% of candidates to be REJECTED or genuinely UNCERTAIN.
Trust the impression. When your overall feeling about something conflicts with your analytical decomposition, investigate the feeling first. The impression is data.
The Flow
U0: GROUND (Identify exemplars -- what does good look like?)
|
U1: MAP (Apply techniques to find the complete possibility space -- number everything)
|
[G1: GENERATE -- for creative domains, produce candidate artifacts]
|
A1: TEST (ARAW top candidates -- number every finding, recurse to bedrock)
|
<- Loop back to U1 if A1 revealed genuinely new directions (max 3 loops)
|
U2: EDGE-CASE (Find where validated candidates break -- number everything)
|
A2: VALIDATE (Test edge cases, produce verdicts with derivation trails)
|
REGISTRY (Compile ALL numbered items from all phases)
|
SYNTHESIS (Derive conclusions ONLY from registry)
When to Skip Steps
U0: Skip for pure logic/math, or when user supplied references
G1: Skip for analytical domains (strategy, engineering, research). Use for design, writing, creative work.
Feedback loop: Skip when A1 produces no surprises
Phase 1: EXPLORATION
U0: Ground in Exemplars
Before analyzing, ask: What does good look like in this domain?
Identify 3-5 best existing examples
Note what they share (likely fundamental) vs differ on (likely stylistic)
Record the felt impression -- this is your perceptual anchor throughout
U1: Map the Space
Apply techniques to find candidates. Number every finding: U1, U2, U3...
[U1] EXPLICIT: [what the input directly states]
[U2] IMPLICIT: [what's assumed but not said]
[U3] PRESUPPOSED: [what must be true for the input to make sense]
[U4] BUNDLED: [separate assertions packed together]
[U5] META: [claims about the type of question or approach]
Then apply techniques (select based on domain):
Technique 1: STATE SPACE -- What states could this be in?
[U6] [alternative 1]
[U7] [alternative 2]
[U8] [the negation]
[U9] [the "do nothing" option]
[U10] [the reframe -- what if the question is wrong?]
Technique 2: INSTANCE-TO-CATEGORY -- What is this an instance of? What siblings?
[U11] [X] is an instance of [CATEGORY]
[U12] Sibling: [sibling 1]
[U13] Sibling: [sibling 2]
Technique 3: PARAMETER VARIATION -- What variables, what ranges?
Artifacts should be concrete enough to evaluate (actual code, actual prose, actual layout)
Don't evaluate yet -- that's A1's job
Unconventional requirement: At least one candidate must be genuinely unconventional -- not just the obvious alternative. If every candidate feels safe, you haven't explored far enough.
A1: Test with ARAW
For each top candidate, build a numbered AR/AW tree. Number every finding: F1, F2, F3...
Candidate [G1 or U-number]:
ASSUME RIGHT:
[F1] If right: [implication] -- Necessary/Probable/Possible
[F2] If F1 right: [deeper implication]
[F3] [-> BEDROCK-TEST: specific test]
[F4] FORECLOSED if right: [what becomes impossible]
ASSUME WRONG:
[F5] Wrong because: [reason] -- Fatal/Serious/Conditional
[F6] If F5 holds: [deeper reason]
[F7] [-> BEDROCK-OBSERVE: observable fact]
[F8] Alternative derived from F5: [what F5 points toward]
[F9] Wrong because: [second reason] -- Fatal/Serious/Conditional
...
Classification:
AR: Necessary / Probable / Possible / Foreclosed
AW: Fatal / Serious / Conditional
Stop ONLY at bedrock: BEDROCK-TEST, BEDROCK-LOGIC, BEDROCK-OBSERVE, BEDROCK-TENSION
Feedback Loop
After A1: did testing reveal genuinely new candidates not in U1? If yes, loop back to U1 to map the expanded space. Number new findings continuing from where you left off. Max 3 loops. Converge when no new candidates emerge.
U2: Find Edge Cases
For each surviving candidate, find where it breaks. Number everything: E1, E2, E3...
[E1] Boundary: [condition where candidate breaks]
[E2] Scale failure: [what happens at 10x/100x]
[E3] Temporal limit: [when does this stop working?]
[E4] Stakeholder conflict: [who disagrees and why]
[E5] Context dependency: [where this only works in specific context]
Also for rejected candidates: under what conditions would they work?
[E6] Rejected [G-number] works if: [specific condition]
A2: Validate Edge Cases
Test each edge case with quick AR/AW. Continue numbering findings.
[E1] "[boundary condition]"
[F30] AR: [why candidate might survive this] -- Necessary/Probable/Possible
[F31] AW: [why this edge case kills it] -- Fatal/Serious/Conditional
[F32] [-> BEDROCK-TEST: specific test]
Phase 2: FINDING REGISTRY
After ALL exploration is complete, compile EVERY numbered item into a categorized registry. Nothing from Phase 1 gets left out.
Verdict values (derived from the tree, not asserted):
VALIDATED: AR evidence reaches bedrock, AW reasons don't reach fatal bedrock, edge cases survived
REJECTED: AW fatal reason reaches bedrock. Record WHY -- rejected candidates may work in other contexts.
DAMAGED: Serious AW reasons found but none individually fatal at bedrock
CONDITIONAL: Wrong under specific conditions, right under others (state both)
UNCERTAIN: Neither side reached bedrock -- needs more investigation
Rules for the registry:
Every U-numbered, G-numbered, F-numbered, and E-numbered item from Phase 1 appears here. No exceptions.
Verdicts must be DERIVED from the tree, not asserted. Point to specific findings.
If a verdict is unclear, mark UNCERTAIN, not VALIDATED.
Phase 3: SYNTHESIS
Derived entirely from the registry. No new findings introduced here.
ORIGINAL INPUT: [restated]
SPACE SIZE: [total unique findings from registry]
WHAT THE ANALYSIS ACTUALLY FOUND:
[Numbered list of EVERY substantive finding, referencing item numbers]
1. [finding, from U-numbers and F-numbers]
2. [finding, from F-numbers]
3. [finding, from E-numbers and F-numbers]
...
KEY TENSIONS:
[Any items that contradict each other. Reference numbers.]
1. [item] vs [item]: [what this tension means] -- TYPE: [resource allocation / information gap / optimization frontier / commitment decision]
2. ...
VOI RANKING (Value of Information -- which findings matter most):
1. [highest-VOI finding -- item number -- learning this changes the most]
2. [second highest -- item number]
3. [third -- item number]
LOAD-BEARING ASSUMPTIONS:
[Assumptions that, if wrong, change everything -- item numbers only]
HIDDEN DIMENSIONS:
[Axes the original input didn't mention but exists on -- item numbers only]
WEAKEST LINKS:
[Which findings are Possible/Conditional rather than Necessary/Fatal?
These are where analysis might break. Reference item numbers.]
ALTERNATIVES DERIVED FROM ANALYSIS:
[Only alternatives that emerged from wrongness reasons. Each cites F-numbers.
If no alternatives emerged, say "None derived -- further exploration needed."]
1. [alternative] -- derived from [F-numbers]
2. ...
TESTABLE PREDICTIONS:
- [prediction derived from specific item numbers]
- [prediction derived from specific item numbers]
DO_FIRST ACTIONS:
1. [action] -- WHO: [Claude/user] -- resolves: [item numbers]
2. [action] -- WHO: [Claude/user] -- resolves: [item numbers]
...
UNRESOLVED:
- [candidates that stayed UNCERTAIN -- what would resolve them]
- [findings that stayed Possible -- what would confirm or deny them]
READY FOR:
- /ar [specific high-VOI claim] -- to go deeper on rightness
- /aw [specific high-VOI claim] -- to go deeper on wrongness
- /u [specific dimension to explore further]
Depth Scaling
Depth
Min Candidates (U1)
Min Edge Cases (U2)
Min ARAW Levels
Min Total Findings
Min Output Lines
1x
5
3
3
20
400
2x
8
5
4
35
800
4x
12
8
5
55
1600
8x
18
12
6
85
3200
16x
25
18
7
130
6400
32x
35
25
8
200
12800
Default depth: 2x. Detect from user input ("uaua 8x" -> 8x). These are FLOORS.
Phase Awareness
UAUA is not equally useful for all phases of creative work:
