Bloom's AI Collaboration Framework (/think) v2

A structured thinking skill that guides human-AI collaboration through Bloom's cognitive levels. The human leads evaluation and judgment (Levels 5-6), AI handles research, execution, and documentation (Levels 1-3), and both collaborate on analysis (Level 4).

Core principle: Always start at Level 5 (Evaluate). Never jump to building, drafting, or coding before the human has answered the evaluative questions.

Interaction model: Use the AskUserQuestion tool for ALL decision points. This makes the framework faster and more guided than free-form Q&A.

Framework Reference

┌─────────────────────────────────────────┐
│  Level 6: Create       HUMAN → AI      │  Direction and vision
│  Level 5: Evaluate     HUMAN → AI      │  Judgment and decisions
│  Level 4: Analyze      HUMAN ↔ AI      │  Research and comparison
├─────────────────────────────────────────┤   ← TRUST BOUNDARY
│  Level 3: Apply        AI → HUMAN      │  Draft, build, produce
│  Level 2: Understand   AI → HUMAN      │  Explain in plain language
│  Level 1: Remember     AI alone        │  Save to Second Brain
└─────────────────────────────────────────┘

L1-L3 sit below the trust boundary — output has an external referent (source post, document, runtime behavior) that can be verified. L4-L6 sit above it — ground truth is the human's judgment and values, which only the human holds. See the Auditability Principle section at the end of this file.

Commands

/think [topic] — Deep mode (default)

Full 6+3 questions + mandatory L4↔L5 spiral + Level 3 deliverable + save. Use for any decision, project, or design effort where the stakes are medium or higher.

/think quick [topic] — Quick mode

3 questions only (Purpose, Success, Scope). No L4 research unless the human asks. No Round 2. Save is optional. Use for low-stakes daily decisions, personal-pillar exploration, fast topic framing, or when you just need a starting point.

/think continue — Resume a session

Resume from where you left off. Check the Second Brain for any saved session state.

/think summary — Show current state

Display which levels have been completed, what decisions were made, and what is next.

/think save — Save session to Second Brain

Save the full session output (decisions, research, deliverables) to the appropriate vault location.

/think feedback [text] — Capture friction

Append a feedback note to ~/.claude/skills/think/feedback.jsonl:

python3 -c "
import json, datetime, sys
note = {
  'ts': datetime.datetime.utcnow().isoformat(),
  'topic': '<current session topic or empty>',
  'text': sys.argv[1]
}
with open('/Users/jneaimi/.claude/skills/think/feedback.jsonl', 'a') as f:
    f.write(json.dumps(note) + '\n')
print('Feedback saved.')
" "<user text>"

No AskUserQuestion needed. Confirm with "Feedback saved." and continue.

/think pillars [days] — Life balance report

Read saved /think sessions from ~/vault/knowledge/decisions/, ~/vault/knowledge/learnings/, and ~/vault/projects/*/decisions/. Parse frontmatter tags: for pillar references. Count sessions per pillar over the last N days (default: 30). Reference pillar definitions from ~/.claude/skills/think/pillars.md (file maintained by a parallel agent). Report format:

Last 30 days — pillars touched:
  Work / Craft         9 sessions
  Decisions            2 sessions
  Identity / Direction 1 session
  Health               0 sessions
  Money                0 sessions
  Relationships        0 sessions
  Learning             0 sessions
  Time / Energy        0 sessions
  Faith                0 sessions

Imbalance: heavy on Work; six pillars untouched.
Suggestion: try /think on Health or Relationships this week.

If pillars.md does not exist yet, use the list above as the default pillar set and note: "Using default pillar list — pillars.md not found."

Level 5: Evaluate (HUMAN LEADS)

This is always the first step. Do not skip it.

Step 1: Detect the Domain

Context-aware skip rule: If the prior 5+ messages of the conversation clearly establish a domain (e.g., the user has been discussing architecture decisions, a specific codebase, or a product spec), skip the Domain detection question. Instead, announce inline:

"Detected domain: Code/Architecture — adjust if wrong, otherwise we will proceed."

Only ask the explicit Domain question when /think is invoked cold (no relevant prior context). Cold invocation: the first message of a session, or a topic that does not follow from what was discussed.

When the Domain question IS needed, use AskUserQuestion:

AskUserQuestion (1 question):

Question: "What type of thinking is this?"
Header: "Domain"
Options (propose the best fit first):
  - [Detected domain] (Recommended) — [brief description]
  - [Second most likely] — [brief description]
  - [Third option] — [brief description]
  (User can type freely to specify)

Domain table:

Step 2: Ask the Universal Questions

#### Quick mode — Round 1 only (3 questions)

AskUserQuestion (3 questions):
  Q1 — PURPOSE, Q3 — SUCCESS, Q5 — SCOPE

Stop after Round 1. Proceed directly to a brief framing summary. No L4 research, no L3 deliverable, no save required. If the human asks to go deeper, switch to deep mode.

#### Deep mode — Round 1 (3 questions)

After domain is confirmed, ask the first three evaluative questions. The AI does NOT propose answers — options are directional framings to help the human think, not AI-proposed solutions. This is the key difference from /prebuild.

AskUserQuestion (3 questions):

Q1 — PURPOSE
  "What is the purpose of this [agreement/product/decision/...]?"
  Options — framing options, not answers:
    - "[Domain-specific framing A]"
    - "[Domain-specific framing B]"
    - "[Domain-specific framing C]"

Q2 — RESOURCES
  "What does each side bring?" / "What do you have vs. need?"
  Options — scope framings:
    - "[Resource framing A]"
    - "[Resource framing B]"

Q3 — SUCCESS
  "What does success look like?"
  Options — outcome framings:
    - "[Outcome A — ambitious]"
    - "[Outcome B — practical]"
    - "[Outcome C — minimal viable]"

#### Deep mode — Round 2: Risk, Scope, Commitment (3 questions)

GATE: Round 2 is mandatory in deep mode. Do not produce a Level 3 deliverable until all 6 questions have answers. Do not skip Round 2 even if Round 1 answers seem sufficient.

AskUserQuestion (3 questions):

Q4 — RISK
  "What are you afraid of?" / "What could go wrong?"
  multiSelect: true
  Options — common risks for this domain (4 options):
    - "[Risk A — most likely]"
    - "[Risk B — highest impact]"
    - "[Risk C — often overlooked]"
    - "[Risk D — domain-specific]"

Q5 — SCOPE
  "What is in vs. out of scope?"
  Options:
    - "[Narrow scope]"
    - "[Medium scope]"
    - "[Broad scope]"

Q6 — COMMITMENT
  "How deep are you going?"
  Options:
    - "[Light — exploration/test]"
    - "[Medium — serious but reversible]"
    - "[Heavy — full commitment / All-in]"

#### Deep mode — Round 3: Epistemic Prompts (mandatory for Heavy commitment)

Round 3 is a single AskUserQuestion call with three epistemic prompts. It is mandatory when Q6 = Heavy or All-in. For Light/Medium commitment it is optional — offer it but do not block progress if the human skips.

AskUserQuestion (3 questions):

Q7 — ASSUMPTIONS
  "What are you assuming to be true about this situation?"
  (free-form — no options; the human writes their assumptions)

Q8 — REVERSIBILITY
  "If this goes wrong, how hard is it to undo?"
  Options:
    - "Fully reversible — low stakes"
    - "Partially reversible with cost"
    - "Difficult to reverse — significant switching cost"
    - "Irreversible — one-way door"

Q9 — FALSIFIABILITY
  "What would change your mind?"
  (free-form — no options; the human writes the falsifying condition)

Step 3: Handle "Argue Mode" — User Writes Specifics Instead of Picking

When the user types their own text instead of selecting an option, accept it as the answer. Do not prompt them to pick from the list again.

Rule: Treat any free-text response to an AskUserQuestion as a valid answer. Acknowledge it and check for downstream implications:

"Got it — you are saying [paraphrase of what they wrote]. Does that change anything else we should account for?"

Then continue to the next question or step.

Example:

• AskUserQuestion offers: "Strategic partnership / One-time transaction / Risk protection"
• Human types: "It is a pilot with option to renew — somewhere between strategic and one-time"
• Response: "Got it — you are saying this is a pilot with renewal option, not a full commitment yet. Does that change how we should frame the exit clause?"

Step 4: Compile and Display Level 5 Summary

After all rounds are complete, compile answers into a clear summary. If any answer was free-text (Argue mode), incorporate it exactly. If any answer is vague or contradicts another, use a follow-up AskUserQuestion to clarify — do not guess.

## Level 5 Summary

- **Purpose:** [compiled answer]
- **Resources:** [compiled answer]
- **Success:** [compiled answer]
- **Risks:** [compiled answer — list all selected]
- **Scope:** [compiled answer]
- **Commitment:** [compiled answer]
- **Assumptions:** [Q7 if answered]
- **Reversibility:** [Q8 if answered]
- **Falsifying condition:** [Q9 if answered]

Moving to Level 4: Analyze...

Level 4: Analyze (COLLABORATION)

Step 5: Research Based on Answers

Based on the human's Level 5 answers, perform relevant research:

• Web search for domain-specific knowledge, legal requirements, market data,

best practices

• Vault search using the brain-search agent to find related Second Brain notes
• Code search if the topic involves existing codebases

Step 6: Present Analysis AND Guarantee a Loop-Back Check

Structure your analysis as:

1. Key findings — what you discovered that is relevant to their answers
2. Risks identified — things the human may not have considered
3. New decisions surfaced — questions that emerged from the research

Mandatory loop-back rule: After presenting the analysis, you MUST check whether any finding introduces a decision the human has not yet made. If yes, return to Level 5 with an AskUserQuestion before proceeding to Level 3. This loop-back is not optional — it must happen at least once per deep mode session.

If the analysis reveals no new decisions (genuinely none), explicitly state: "Level 4 complete — no new decisions surfaced. Proceeding to Level 3."

AskUserQuestion (1 question) — loop-back example:

Question: "[Specific new decision that emerged from analysis]"
Header: "New decision"
Options:
  - "[Option A — based on research finding]"
  - "[Option B — alternative based on research]"
  - "[Option C — if applicable]"

The L4↔L5 loop may repeat. Each loop produces a cleaner Level 3 deliverable.

Signs that a loop-back is required:

• Research revealed a choice the human has not made
• There is a fork in approach that depends on human preference
• A risk was found that changes the scope or commitment answer
• The human's initial answers assumed something the research contradicts

Level 3: Apply (AI LEADS, HUMAN VERIFIES)

Step 7: Gate Check Before Producing the Deliverable

Before producing any Level 3 output, verify:

• [ ] All 6 questions answered (deep mode) or all 3 questions (quick mode)
• [ ] At least one L4↔L5 loop completed OR explicitly stated no new decisions
• [ ] Q9 Round 3 answered if commitment was Heavy/All-in

If any gate is unmet, do not produce the deliverable. Return to the missing step.

Step 8: Produce the Deliverable

Match the format to the domain:

Integration with Other Skills

• Market data needed → use /research collection pipeline
• Vault save → use /brain patterns and templates
• Code implementation → use /prebuild for the build phase
• Architecture diagrams → use /diagram

Level 2: Understand (AI EXPLAINS)

Step 9: Explain What Was Produced

After delivering the output:

• What each section/component does and why it is there
• What the human is committing to or agreeing to
• What they should pay attention to or verify
• What the limitations are (what this does NOT cover)

Tailor depth to expertise:

• Expert: Brief highlights of non-obvious choices only
• Some familiarity: Section-by-section walkthrough
• Beginner: Full plain-language explainer

Level 1: Remember (AI SAVES)

Step 10: Save to Second Brain

In deep mode, offer a save prompt. In quick mode, save is optional — offer but do not block.

AskUserQuestion (1 question):

Question: "Where should we save this session's output?"
Header: "Save to"
Options:
  - "[Most likely vault location based on domain]" (Recommended)
  - "[Alternative location]"
  - "Don't save — this was exploratory"

Save location by type:

Always:

• Add proper frontmatter (type, created, tags — include pillar tag if relevant)
• Use YYYY-MM-DD date format
• Update the folder's index.md
• Cross-reference related vault notes

Domain-Adapted Option Examples

Legal/Agreement:

• Purpose: "Strategic partnership" / "One-time transaction" / "Risk protection (NDA/non-compete)"
• Resources: "Both sides contribute equally" / "One side brings capital, other brings expertise" / "Primarily one-sided obligation"
• Success: "Long-term relationship with clear governance" / "Clean transaction with no loose ends" / "Protective boundary that holds under pressure"
• Risks: "Scope creep beyond original terms" / "Exit clause too weak or too restrictive" / "IP ownership ambiguity" / "One party not delivering"
• Scope: "Single project or transaction" / "Ongoing relationship with review periods" / "Umbrella agreement covering multiple workstreams"
• Commitment: "Non-binding exploration (MOU)" / "Binding but limited scope" / "Full legal commitment"

Product/Feature:

• Purpose: "Solve a specific user pain point" / "Expand into a new market/audience" / "Reduce operational cost internally"
• Resources: "Solo build with existing tools" / "Small team, some skills missing" / "Full team, needs budget approval"
• Success: "Users adopt without being told" / "Measurable metric improvement (define which)" / "Internal team saves X hours/week"
• Risks: "Nobody uses it" / "Technical debt outweighs value" / "Scope balloons past resources" / "Security/compliance gap"
• Scope: "MVP — one core flow" / "Full feature — happy path + edge cases" / "Platform — extensible for future needs"
• Commitment: "Side project / experiment" / "Committed feature with deadline" / "Core product bet"

Decision:

• Purpose: "Choose between known options" / "Decide whether to proceed at all" / "Prioritize competing demands"
• Resources: "Have all info needed" / "Missing key data — need research first" / "Have opinions but no consensus"
• Success: "Clear decision with documented reasoning" / "Alignment across stakeholders" / "Reversible choice with defined exit criteria"
• Risks: "Analysis paralysis" / "Picking the default without examining it" / "Irreversibility — can't undo easily" / "Missing a better option not yet considered"
• Scope: "This decision only" / "This decision + downstream implications" / "Framework for recurring decisions of this type"
• Commitment: "Quick call — low stakes" / "Significant — affects next 3-6 months" / "Strategic — affects direction for 1+ year"

Code/Architecture:

• Purpose: "New system from scratch" / "Extend existing system" / "Replace/migrate legacy"
• Resources: "Existing codebase + team conventions" / "Greenfield — no constraints" / "Constrained by existing infra/contracts"
• Success: "Passes tests + code review" / "Handles production load + edge cases" / "Maintainable by team in 6 months"
• Risks: "Breaking existing functionality" / "Over-engineering for current needs" / "Security vulnerability" / "Performance regression"
• Scope: "Single component or endpoint" / "Full feature across multiple files" / "System-level architecture change"
• Commitment: "Prototype / spike" / "Production feature" / "Long-term architecture bet"

Content/Writing:

• Purpose: "Thought leadership — establish authority" / "Education — teach something specific" / "Engagement — start a conversation"
• Resources: "Original research / data" / "Personal experience / opinion" / "Curated from existing sources"
• Success: "Reader takes a specific action" / "Reader changes how they think about X" / "Reader shares it"
• Risks: "Too generic — says nothing new" / "Too niche — audience too small" / "Tone mismatch for platform" / "Factual error damages credibility"
• Scope: "Single post" / "Article or thread" / "Series or campaign"
• Commitment: "One-off piece" / "Part of a content strategy" / "Flagship content for the brand"

Business/Strategy:

• Purpose: "Pursue a new revenue opportunity" / "Defend or improve current position" / "Explore — not committing yet"
• Resources: "Existing team and skills" / "Need to hire or partner" / "Need funding first"
• Success: "Revenue target in X months" / "Market position shift" / "Validated learning (even if we pivot)"
• Risks: "Market doesn't exist or is too small" / "Execution — can't build fast enough" / "Competition — someone else gets there first" / "Cash runway runs out"
• Scope: "One market, one product" / "One market, multiple offerings" / "Multiple markets"
• Commitment: "Bootstrapping / lean" / "Raising seed" / "All-in bet"

Session State Tracking

Throughout the session, maintain awareness of which levels have been completed. When asked for a summary (/think summary), display:

## /think Session: [Topic]
Mode: [Deep / Quick]

### Level 5: Evaluate [status]
- Purpose: [one-line summary]
- Resources: [one-line summary]
- Success: [one-line summary]
- Risk: [one-line summary]
- Scope: [one-line summary]
- Commitment: [one-line summary]
- Assumptions / Reversibility / Falsifiability: [if answered]

### Level 4: Analyze [status] ([N] loops)
- Loop 1: [what was researched, what decisions emerged]
- Loop 2: [follow-up research, final decisions]

### Level 3: Apply [status]
- Deliverable: [what was produced]

### Level 2: Understand [status]
- Explanation: [delivered / skipped — human is expert]

### Level 1: Remember [status]
- Status: [not yet saved / saved to path]

Rules

1. NEVER skip Level 5. Even if the human says "just build it" — ask at

least Q1, Q3, Q5 (purpose, success, scope) via AskUserQuestion before proceeding.

1. NEVER answer Level 5 questions for the human. Offer directional options

to help them think; they provide the substance.

1. ALWAYS use AskUserQuestion for domain detection, the universal

questions, Level 4 loop-backs, and Level 1 save decisions.

1. GATE: Never produce a Level 3 deliverable without completing all

required rounds (all 6 in deep mode) and at least one L4↔L5 loop-back check.

1. MANDATORY loop-back: After every Level 4 analysis, explicitly check for

new decisions. If found, return to Level 5 before proceeding. If none, state it explicitly.

1. Accept Argue mode answers. When the user types free text instead of

picking an option, treat it as a valid answer. Paraphrase and continue.

1. Context-aware domain skip. When 5+ prior messages establish a domain,

announce the detected domain inline and skip the Domain detection question.

1. Adapt question wording AND options to the domain — do not use generic

versions when domain-specific ones are clearer.

1. Keep the human aware of which level you are operating at. Use phrases

like "Now moving to Level 4" or "This is a Level 5 question — only you can decide."

1. Use previews on AskUserQuestion when comparing concrete alternatives.

Skip previews for abstract/directional questions.

1. Auto-fire transparency. When /think is invoked by Claude (Skill tool

auto-fired because context matched) rather than typed by the user, announce it on the first turn: "Auto-firing /think because [reason]. Reply 'no skip' to disable for this session." Auto-fire is often the right call, but silent auto-fire violates the framework's first principle — the human must know they are entering Level 5. Historical data: ~40% of /think invocations (10 of 25 across 2026-03-16 → 2026-04-25) were auto-fires. Make activation visible.

Auditability Principle

L1-L3 output sits below the trust boundary because it has an external referent: the source document exists, the code runs or fails, the cited fact can be checked. Verification is possible.

L4-L6 output sits above the trust boundary because the ground truth is the human's judgment and values. No external source can confirm that the right trade-off was made, that the risk priority matches actual stakes, or that the direction aligns with what the human cares about. Only the human holds that ground truth.

This is the framework's central reframe: the AI is not "helping with the hard parts" above the boundary — it is assisting a process whose correctness only the human can validate. When the human rubber-stamps L4-L6 output without genuine evaluation, the collaboration fails even if the output looks sophisticated.

Failure Modes

These are known ways /think breaks down:

(a) Human cannot competently audit Level 3 output. If the deliverable is in a domain where the human has no ability to evaluate correctness (e.g., a legal document in an unfamiliar jurisdiction), L3 output may pass unverified. Mitigation: the Level 2 explanation step exists specifically for this — always run it for unfamiliar domains.

(b) AI confidently wrong at L1, source not checked. The AI recalls training data with apparent confidence whether it is correct or not. If a recalled fact is not verified against a primary source, the entire downstream chain rests on a potentially incorrect foundation. Mitigation: flag any L1 claim that could be jurisdiction-specific, version-specific, or rapidly-changing.

(c) L5 questions answered superficially. When the human picks the first option quickly without genuine reflection, Level 5 becomes a checkbox rather than an evaluation. The output will then reflect the path of least resistance, not the human's actual goals. Mitigation: if answers come back in under 10 seconds for all six questions, surface one follow-up before proceeding.

(d) L4 skipped because output looks "good enough." If the Level 5 summary already sounds complete, it is tempting to jump straight to Level 3. The L4 research step frequently surfaces contradictions or missing information that would otherwise produce a flawed deliverable. Skipping it is the most common source of rework. Mitigation: the Level 3 gate enforces this structurally.

(e) Downstream agents drift from /think-derived decisions. When a /think session produces a decision that is passed to another agent (architect, developer, content drafter), the agent may receive only the deliverable and not the evaluation context behind it. Subsequent choices by that agent may silently contradict the L5 decisions. Mitigation: always save the full Level 5 Summary (not just the deliverable) and link it in the handoff note to downstream agents.