Published March 1, 2026
A Claude Code plugin that improves conversation quality before execution begins. Conversations go wrong in predictable ways — unclear goals, untested assumptions, unexplored options, decisions made without considering what might go wrong. Good Thinking is a set of thinking skills designed to catch these failure modes early.
Each skill follows a two-part pattern: an evaluation step that determines whether it's relevant, then a multi-step workflow that runs only after the user confirms. Skills are discovered eagerly — Claude suggests them when they might apply, rather than waiting for explicit invocation.
Context-blind agents
Skills spawn subagents without conversation context, ensuring fresh perspectives that aren't anchored to the user's existing framing
Friction preservation
Dissent and disagreement aren't smoothed away by summarization — conflicting views are presented as-is for the user to resolve
Eager discovery
Skills are loaded when they might apply, letting the evaluation step decide relevance rather than relying on prescriptive triggers
Human-in-the-loop
Every skill suggests itself conversationally and waits for confirmation — nothing runs without explicit consent
Sharpen the problem before solving it
Sharpens problem and goal statements
Creates an initial framing of the user's intent, then spawns a context-blind agent to evaluate it for precision, clarity, completeness, and scope. If gaps are found, it asks targeted clarifying questions. A second agent then checks whether the problem could be better framed at a different level of abstraction — surfacing alternative framings the user may not have considered.
Breaks problems into independent parts
Proposes an initial decomposition, then iteratively validates it against MECE criteria using a context-blind auditor. The auditor checks that parts are mutually exclusive, collectively exhaustive, at uniform abstraction levels, and actionable. The loop continues until all criteria pass — preventing the common failure of breakdowns that look clean but overlap or miss dimensions.
Tests hypotheses under adversarial scrutiny
Spawns two parallel context-blind agents: an Analyst who builds the strongest case FOR the thesis (extracting supporting evidence and validating patterns) and a Skeptic who builds the case AGAINST (finding counter-evidence, alternative interpretations, and logical gaps). Each then sees the other's findings and refines their position. The result is a defensibility spectrum — solid ground, contested claims, and unsupported claims — rather than a binary verdict.
Stress-test decisions before committing
Surfaces concerns the user hasn't considered
Identifies the decision as a clear proposition, then spawns a context-blind agent to generate challenges — hidden assumptions, practical risks, opportunity costs, and second-order effects. Results are filtered through relevance, novelty, importance, and actionability gates. Only concerns that survive all four filters are presented. If nothing survives, the user is told the decision was stress-tested with no new issues found.
Checks whether options are well-constructed before evaluation
Before the user evaluates a set of options, a context-blind agent tests them against seven framing integrity criteria: missing options, dimensional collapse, description asymmetry, criteria rigging, false exclusivity, scope mismatch, and constraint fabrication. Marginal or checklist-filling concerns are calibrated out. If real issues survive, they're raised before evaluation begins — not after a choice has already been made.
Generate and organize ideas with diversity built in
Generates diverse ideas using parallel perspective agents
Frames the problem as "How Might We" questions, then selects 2–4 diversity primes — different lenses or contextual variations designed to push thinking in different directions. Each prime gets its own parallel context-blind agent that generates 8–10 ideas from that perspective. Results are automatically consolidated using the grouping skill below.
Consolidates raw ideas into strategic themes
Takes a raw list of ideas, deduplicates across sources, clusters similar items, and creates 2–7 strategic themes at a uniform abstraction level. Validates that themes are MECE, actionable, and collectively complete. Used automatically after brainstorming, or standalone when ideas accumulate organically in conversation.
1. Open the plugin manager in Claude Code
/plugin2. Add the Extreme Clarity marketplace
extremeclarity/claude-plugins3. Install the goodthinking plugin from the marketplace