it-i-think

name	it - I Think
description	Process 'I think' statements into explicit claims with confidence levels, evidence/assumption separation, and next verification actions.
output	{"format":"prose"}

IT - I Think

Input: $ARGUMENTS

Interpretations

Before executing, identify which interpretation matches the user's input:

Interpretation 1 — Decompose a belief into testable parts: The user has an "I think" statement and wants it unbundled into its core claim, evidence, assumptions, confidence level, and recommended next action. Interpretation 2 — Calibrate confidence on an uncertain claim: The user is primarily unsure how confident they should be about something and wants help distinguishing what they know from what they assume. Interpretation 3 — Route a vague intuition to the right analysis: The user has a gut feeling or loose opinion and doesn't know what kind of thinking it needs — factual verification, strategic stress-testing, value examination, or something else.

If ambiguous, ask: "I can help with decomposing a belief into testable parts, calibrating your confidence level, or figuring out what kind of analysis your intuition needs — which fits?" If clear from context, proceed with the matching interpretation.

---

Core Principles

1. "I think" hides structure. Every "I think" statement bundles a claim, evidence (or lack of it), assumptions, and a confidence level. Unbundling these is the skill's core operation.

2. Confidence is not binary. People say "I think" for claims they're 20% sure about and claims they're 90% sure about. The same words, vastly different states. Calibrating confidence before acting prevents both recklessness and paralysis.

3. Claims are often compound. "I think we should restructure the team" contains at least three claims: the team has a structural problem, restructuring would fix it, and now is the right time. Decompose before testing.

4. Evidence and assumptions look alike. "I think this because we tried it before and it failed" sounds like evidence but might be an assumption — was the context the same? Did it fail for the reason assumed? Separate what's observed from what's inferred.

5. The right next action depends on claim type AND confidence. A low-confidence factual claim needs verification. A high-confidence strategic claim needs stress-testing. A medium-confidence ethical claim needs value examination. One routing table doesn't fit all.

---

Phase 1: Claim Extraction

[I1] RAW_STATEMENT: [the user's "I think" statement, quoted]
[I2] CORE_CLAIM: [the central claim, stated neutrally]
[I3] CLAIM_TYPE: [factual | strategic | evaluative | predictive | normative | preference]

Compound Claim Check

[I4] IS_COMPOUND: [yes/no]
[I5] SUB-CLAIMS (if compound):
  [I5a] [sub-claim 1]
  [I5b] [sub-claim 2]
  [I5c] [sub-claim 3]

Claim Type	Example	Testing Method
Factual	"I think the server is down"	Check — verify against reality
Strategic	"I think we should pivot"	Stress-test — AR/AW analysis
Evaluative	"I think this code is bad"	Criteria — what does "bad" mean?
Predictive	"I think this will fail"	Forecast — what evidence supports/refutes?
Normative	"I think we should be more transparent"	Values — whose values? What tradeoffs?
Preference	"I think React is better"	Criteria — better for what? By what measure?

---

Phase 2: Evidence/Assumption Separation

For each claim (or sub-claim if compound):

[I6] EVIDENCE (observed, testable):
  [I6a] [evidence 1] — SOURCE: [where this comes from]
  [I6b] [evidence 2] — SOURCE: [source]

[I7] ASSUMPTIONS (inferred, untested):
  [I7a] [assumption 1] — TESTABLE: [yes/no] — TEST: [how to test if yes]
  [I7b] [assumption 2] — TESTABLE: [yes/no] — TEST: [how]

[I8] GAPS (neither evidence nor assumption — just missing):
  [I8a] [what's unknown] — MATTERS: [high/medium/low] — FINDABLE: [yes/no]

Separation Test

For each piece of "evidence," ask:

• Did I directly observe this, or am I inferring it? (Inference → assumption)
• Could someone else verify this independently? (No → assumption)
• Am I using a past experience as evidence for a different situation? (Probably → assumption)

---

Phase 3: Confidence Calibration

[I9] STATED_CONFIDENCE: [what the user seems to feel — from tone/hedging]
[I10] CALIBRATED_CONFIDENCE: [after evidence/assumption analysis]
  LEVEL: [very low (<20%) | low (20-40%) | medium (40-60%) | high (60-80%) | very high (>80%)]
  REASONING: [what drives this level]

[I11] CONFIDENCE_DRIVERS:
  UPWARD: [what makes confidence higher — e.g., strong evidence, domain expertise]
  DOWNWARD: [what makes confidence lower — e.g., untested assumptions, novel situation]

Common Miscalibrations

Pattern	What Happens	Correction
Expertise inflation	"I've done this before" → overconfidence	Was the context the same?
Hedging as signal	"I think maybe perhaps" → very low stated	Might actually be medium — hedging is social, not epistemic
Certainty anchoring	First impression calcifies	What would change your mind?
Availability bias	Recent vivid example dominates	Is this representative or memorable?

---

Phase 4: Next Action Routing

Based on claim type + calibrated confidence:

Claim Type	Confidence	Route
Factual	Any	→ /ver to verify against evidence
Strategic	High	→ /aw to stress-test (high confidence needs adversarial pressure)
Strategic	Low/Medium	→ /ar to explore what follows if right
Evaluative	Any	→ /evaluate with explicit criteria
Predictive	Any	→ /ht to formulate testable hypothesis
Normative	Any	→ /ve to examine underlying values
Preference	Any	→ /cmp to compare against alternatives with criteria
Compound	Any	→ Decompose first, route each sub-claim separately

[I12] RECOMMENDED_ACTION: /skill-id — [why this is the right next step]
[I13] INVOCATION: /skill-id [specific arguments derived from the claim]
[I14] ALTERNATIVE_IF_WRONG: /skill-id — [backup if first choice doesn't resolve]

---

Phase 5: Output

"I THINK" DECOMPOSITION
========================

ORIGINAL: [quoted statement]

CLAIM: [core claim, stated neutrally]
TYPE: [factual | strategic | evaluative | predictive | normative | preference]
COMPOUND: [yes/no — if yes, sub-claims listed]

EVIDENCE:
- [evidence with source]

ASSUMPTIONS:
- [assumption with testability]

GAPS:
- [unknown with importance]

CONFIDENCE: [level with percentage range]
REASONING: [what drives the confidence level]

→ INVOKE: /skill-id [specific invocation]
  WHY: [why this is the right next step for this type of claim at this confidence]
  IF_WRONG: /skill-id [backup route]

---

Failure Modes

Failure	Signal	Fix
Claim accepted at face value	No evidence/assumption separation	Always unbundle — every claim has hidden structure
Compound claim treated as atomic	Single confidence level for multi-part claim	Decompose compound claims; rate each sub-claim
Confidence not calibrated	Using stated confidence without checking	Apply miscalibration checks
Wrong routing	Strategic claim routed like factual claim	Route by claim TYPE, not just confidence
Evidence confused with assumption	Inference treated as observation	Apply separation test to each piece
Generic routing	Everything goes to /claim	Different claim types need different skills

---

Depth Scaling

Depth	Sub-Claims Checked	Evidence/Assumption Items	Calibration Checks	Routing Alternatives
1x	1	2	1	1
2x	3	4	3	2
4x	5	8	5	3
8x	All	12	All patterns checked	Full routing analysis

Default: 2x. These are floors.

---

Pre-Completion Checklist

• Core claim extracted and stated neutrally
• Claim type classified
• Compound claims decomposed into sub-claims
• Evidence separated from assumptions with sources
• Each assumption tested for testability
• Confidence calibrated (not just using stated confidence)
• Miscalibration patterns checked
• Routing matches claim type + confidence level
• Invocation includes specific arguments (not generic)

---

Integration

• Use from: natural language processing of user input
• Routes to: /ver, /aw, /ar, /evaluate, /ht, /ve, /cmp depending on claim type
• Complementary: /aex (extract hidden assumptions), /nsa (when confidence is very low)
• Differs from /claim: claim does full truth-testing; it decomposes and routes
• Differs from /nsa: nsa handles uncertainty; it handles all "I think" statements including confident ones