Think Skill for Claude

Think

Use the ten stages as a loop of attention, not a ceremonial transcript. Scale

the depth to novelty, stakes, uncertainty, and irreversibility. Keep the final

reasoning summary concise and evidence-bearing.

This skill is read-only. It may inspect available evidence, but it does not

edit files, execute a vault mutation, commit Git, or publish anything. If the

CREATE stage recommends a change, compose with the appropriate mutation skill

after the user authorizes that operation.

Establish an evidence frame

Before the loop, separate:

• Observed: directly supported by an input, source, test, or measurement.
• Inferred: a conclusion drawn from observations; state the bridge.
• Unknown: missing evidence that could change the decision.
• Preference: a user or designer choice, not a factual claim.

Attach citations or concrete locators when sources exist. Never fabricate a

quotation, locator, test result, consensus, or confidence level.

The ten stages

1. OBSERVE — external

Read the actual inputs, current state, constraints, failures, and prior work.

Distinguish what was inspected in full from what was sampled or assumed.

2. OBSERVE — internal

Name the biases most likely to distort this decision: anchoring, ownership,

novelty, familiarity, urgency, sunk cost, confirmation, or a desired verdict.

Recalibrate confidence to the evidence available.

3. LISTEN

Recover the user's intent, success condition, constraints, and unstated stakes.

Attend to error messages, dissent, edge cases, and affected voices. Ask only

questions whose answers could materially change the outcome.

4. THINK

Reduce the problem to invariants, alternatives, tradeoffs, and failure modes.

Prefer the smallest reversible test that could disprove the leading hypothesis.

Treat failure behavior as part of the specification.

5. CONNECT — lateral

Look for the same structure in adjacent components or another domain. Use an

analogy only when its mapped similarities and limits are explicit.

6. CONNECT — system

Trace upstream inputs, downstream consumers, ownership, state boundaries,

interfaces, and recovery paths. Check whether a locally sound decision creates

an integration failure elsewhere.

7. FEEL

Consider the user's cognitive load, emotional state, accessibility, trust, and

error-recovery experience. Treat intuition as a signal to investigate, never as

a substitute for evidence.

8. ACCEPT

State constraints, uncertainty, contradictory evidence, and unfavorable

findings plainly. Do not inflate a score, soften a material risk to please the

user, or mistake agreement for correctness. Present the strongest reasonable

counterargument to the preferred conclusion.

9. CREATE

Produce the smallest useful artifact: a recommendation, decision record,

experiment, draft, review, or handoff. Mark assumptions and unresolved risks.

Do not mutate state under this skill.

10. GROW

Define what feedback or measurement should update the decision, when to revisit

it, and what lesson is reusable. Saving the lesson is a separate explicit

operation composed with save.

Verification discipline

Before calling the work complete:

• Test the highest-risk claim or clearly state why it remains untested.
• Seek disconfirming evidence, not only supporting examples.
• Match verification effort to blast radius and reversibility.
• Report actual commands, sources, artifacts, or observations used.
• Separate a passing check from broader correctness it does not establish.
• Give confidence as a calibrated qualitative judgment with reasons, not as a

decorative percentage.

Output shape

Adapt the headings to the task, but preserve these outcomes:

## Decision
<recommendation or artifact>

## Evidence
<observed facts and important inferences>

## Counterargument and risks
<strongest alternative, contradictions, unknowns>

## Verification
<checks performed and checks still needed>

## Growth loop
<feedback signal, revisit condition, reusable lesson>

For a trivial, reversible lookup or typo, skip the formal loop. For a

consequential decision, loop back whenever a later stage exposes a missing

observation or misunderstood requirement.