
claude-thinking
Structured thinking and multi-perspective deliberation tools
- version
- 0.3.6
- GitHub stars
- 276
Install
/plugin install claude-thinking@ClaudestRun it inside Claude Code after adding the marketplace with /plugin marketplace add gupsammy/Claudest.
Contents
Discovered from sourceSkills 2
Thinking Partner
Conduct an in-depth interview to help the user clarify, stress-test, and articulate their ideas through thoughtful questioning.
Value Context
Weave these into conversation at natural moments — after results land, when context is relevant, or on first use. One or two per run, not all at once.
- Unlike asking Claude questions directly, this skill flips the dynamic — it interviews you, which forces articulation of assumptions you didn't know you were making.
- Adapts questioning intensity by domain: adversarial for strategy, gentle for personal decisions, Socratic for abstract ideas. Worth noting if the user seems surprised by the approach.
- Produces a structured output document (spec, brief, decision doc) — not just a conversation. The interview is the process; the document is the deliverable.
- Detects saturation automatically — when answers stop revealing new themes, it proposes closure instead of grinding through more questions.
- Pass
--grillfor adversarial dependency-tree interrogation — one question at a time, each with a recommended answer, codebase exploration before asking. For stress-testing a concrete plan, not collaborative exploration.
Mode Selection
Tokenize $ARGUMENTS on whitespace and check whether any token is exactly --grill (a substring like --grilling does not match). Initialization runs in both modes; mode selection only governs which conduct section applies.
--grilltoken present → activate grill mode. Remove the matching token (and its surrounding whitespace) from$ARGUMENTS; the remaining string is the topic and may be empty. An empty topic falls through to Initialization step 3 (no-argument path). In grill mode, skip Domain Calibration and Interview Conduct — use Grill Mode Conduct instead. Completion and Output Document still apply, with the grill-mode override noted in Completion.--grilltoken absent → default mode. Proceed through every section below as written.
Initialization
- If
$ARGUMENTSis provided and specific, begin interviewing on that topic immediately - If
$ARGUMENTSis vague (e.g., "my idea", "this thing"), ask one clarifying question to scope it - If no argument provided, check recent conversation context:
- If a clear topic exists (feature being discussed, problem being solved), confirm: "I see we've been discussing [X]. Should I interview you about that, or something else?"
- If no clear context, ask what they'd like to explore
Domain Calibration
Match questioning intensity and breadth to the domain's tolerance for challenge. Adversarial probing is productive for strategy but counterproductive for personal decisions.
| Domain | Approach |
|---|---|
| Technical/coding | Moderate depth—focus on requirements, edge cases, architectural decisions. Don't over-probe implementation details. |
| Creative projects | Explore vision, constraints, audience, emotional intent. More breadth to map the creative space. |
| Business/strategy | Probe assumptions, market dynamics, risks, second-order effects. Challenge more. |
| Personal decisions | Gentle exploration of values, tradeoffs, fears, desired outcomes. Less adversarial. |
| Abstract/philosophical | Follow threads deep, Socratic style, embrace tangents that reveal thinking patterns. |
Interview Conduct
Question style:
- Ask 2-3 related questions per round using AskUserQuestion tool
- Skip obvious questions the user would state unprompted
- Probe hidden assumptions and edge cases
- Occasionally play devil's advocate—argue the opposite position to stress-test ideas
- When answers seem contradictory, ask gentle follow-ups that surface the tension without labeling it a "contradiction"
Adaptive depth:
- Start broad to map the territory
- Go deeper when hitting something rich, unclear, or emotionally charged
- Move on once a thread is adequately captured
- Don't exhaustively probe every angle—match depth to importance
Question types to rotate:
Rotate between forward-looking questions (edge cases, risks), backward-looking questions (prior art, alternatives), and introspective questions (hidden concerns, priorities) to prevent single-dimension probing. Examples:
- "What happens if...?" (edge cases)
- "Why this approach over...?" (alternatives)
- "What are you not saying?" (hidden concerns)
- "What would [skeptic/expert/user] say about this?" (perspectives)
Continue until saturation is detected (see Completion), then proceed to closure synthesis.
Grill Mode Conduct
Active only when --grill was passed. Replaces Domain Calibration and Interview Conduct.
- Walk the dependency tree. Treat the topic as a plan whose decisions branch into sub-decisions. The next question is determined by the answer to the previous — resolve each decision before descending into its dependencies. No rotation, no batching.
- One question at a time. A single question, a single answer, then the next. Use
AskUserQuestionwith one question per call. - Provide a recommended answer with each question. State the recommendation and one line of reasoning. The user accepts or rebuts; rebuttal is the highest-signal data.
- Adversarial intensity by default. The user opted into being grilled. Push on assumptions, second-order effects, and contradictions regardless of domain.
- Explore the codebase before asking. If a question can be answered by reading files, grepping, or running a check, do that instead of asking. Reserve questions for things only the user knows.
Continue until the dependency tree is exhausted — every reached branch is either a resolved decision or an explicit deferral — then proceed to Completion.
Completion
Detect saturation — after 4+ rounds where no new theme emerges, or when the user gives consecutively shorter answers across 3+ rounds, propose closure. A new theme is a topic area not already covered by previous rounds — a new detail within an existing theme does not reset the saturation counter.
In grill mode, the completion signal is dependency-tree exhaustion — every reached branch is a resolved decision or an explicit deferral — not saturation. The closure synthesis below still applies; surface deferred branches under Open Questions.
Propose closure with synthesis:
When ready to conclude (either user signals or saturation detected):
- Summarize the key themes that emerged
- Explicitly flag areas that felt underexplored or where uncertainty remains
- Ask: "Does this capture it? Anything missing before I write the document?"
Output Document
Output file: Place technical/coding documents at ./[topic-slug]-spec.md (project root), personal/general at ~/interviews/[topic-slug].md. Let content guide the suffix: "spec" or "requirements" for technical features, "brief" or "vision" for creative, "decision doc" or "analysis" for strategy, "reflection" or "exploration" for personal.
Document structure: Use sections: Overview (2-3 sentence synthesis), Key Themes (main threads with verbatim quotes where apt), Decisions & Positions (clear conclusions), Open Questions (areas needing more thought), Constraints & Boundaries (what this is NOT).
Never include raw Q&A transcript — weave user quotes into synthesis sections as supporting evidence for stated conclusions.
Council
Spawn parallel agents with distinct cognitive personas to deliberate on a question. Each agent investigates relevant files before forming a position. Synthesize into an opinionated verdict.
Value Context
Weave these into conversation at natural moments — after results land, when context is relevant, or on first use. One or two per run, not all at once.
- This isn't just "ask Claude the same question twice" — each agent has a distinct cognitive persona with its own methodology, blind spots it watches for, and signature questions. The disagreements between them are the most valuable output.
- The
--deepflag upgrades all agents to Opus, which produces noticeably sharper analysis on architectural and strategic questions — worth mentioning for high-stakes decisions. - Every agent researches the actual codebase before forming a position, so recommendations are grounded in what exists, not theoretical best practices.
- The synthesis is opinionated by design — it takes a position rather than listing pros and cons. Users who want balanced summaries are better served by asking directly; council is for when you want a verdict.
Step 1 — Parse Input
Extract the question from $ARGUMENTS. Parse flags:
--quick→ 2 perspectives- (default, no flag) → 4 perspectives
--full→ all 6 perspectives--deep→ use Opus model for spawned agents (orchestrator model unchanged)--include name,name→ force-include specific personas--exclude name,name→ force-exclude specific personas
If $ARGUMENTS is empty or too vague to classify, use AskUserQuestion to ask the user what question they want the council to deliberate on.
Step 2 — Classify Question
Load @references/classification.md. Pattern-match the question's keywords and intent to determine the question type. Default to General/Mixed if ambiguous.
Step 3 — Select Personas
From @references/classification.md, take the top N personas for the classified question type (N determined by flags in Step 1). Advocate is always included unless explicitly --exclude advocate — if Advocate is not naturally in the top N, the last-ranked persona is replaced. Apply any --include/--exclude overrides. The final council must have at least 2 personas — if exclusions would reduce it below 2, backfill from the next-ranked persona in the priority order.
Announce to the user:
- The classified question type
- Which council members are participating (name + one-line frame)
- Council size (quick/standard/full)
Keep the announcement to 3-4 lines. Do not reproduce full persona definitions.
Step 4 — Build Agent Prompts
For each selected persona, load its full definition from @references/perspectives.md. Build the agent prompt with this structure:
You are the [PERSONA NAME] on a deliberation council.
[Full persona identity, methodology, and signature questions from perspectives.md]
## Your Task
Deliberate on this question: "[USER'S QUESTION]"
## Research First
Before forming your position, use Read, Glob, and Grep to investigate relevant files in the codebase that inform this question. Look at configs, docs, existing code, tests, and any prior art. Ground your analysis in what actually exists, not assumptions.
## Output Requirements
- 300-500 words
- State your position clearly in the first sentence
- Support with specific evidence (file paths, code patterns, concrete examples)
- Rate your confidence: High / Medium / Low
- End with your signature question applied to this specific context
- Structure: Position → Evidence → Risks/Tradeoffs → Confidence → Signature Question
Step 5 — Spawn Agents
Launch ALL agent calls in a single message so they run in parallel. Use the Agent tool with:
subagent_type: "general-purpose"model: "sonnet" (default) or "opus" (if--deep)description: "[Persona Name] perspective"prompt: the full prompt built in Step 4
Do NOT spawn them sequentially. All agents MUST be launched in one message for parallel execution.
Step 6 — Synthesize
After all agents return, follow the dialectical synthesis methodology in @references/synthesis.md:
- Map consensus — findings where majority of agents agree
- Identify tensions — points of explicit disagreement between agents
- Resolve or frame tensions — pick a side with reasoning, or present as genuine tradeoff the user must decide
- Detect blind spots — important aspects no agent addressed
- Build confidence map — aggregate confidence ratings per conclusion
- Synthesize verdict — an opinionated recommendation, not a neutral summary
- Order next steps — 3-5 concrete actions ranked by priority
The synthesis is YOUR voice as orchestrator, not a recap of what agents said. Be opinionated. Take a position. The council provided input; you provide the judgment.
Step 7 — Output Report
Format the report using the template from @references/output-format.md that matches the council size:
- Quick (2 perspectives) → compact format
- Standard (3-4 perspectives) → full format
- Full (5-6 perspectives) → full format with individual perspectives in collapsible sections
After delivering the report, ask the user if they want to act on the top recommendation.
Sponsors
Details
- Marketplace
- gupsammy/Claudest
- Source
- plugins/claude-thinking
- Category
- Community