Decision Framework Skill

2026-07-03 · page · source ↗ · source summary

A Claude Code skill by Fero Volar that transforms Jason Fried's 37signals decision-making guide — 38 questions used internally as mental models — into an adaptive, conversational coaching experience. Rather than dumping a checklist, it asks only the questions that meaningfully improve the current decision, working through four structured phases (Understand → Explore → Reflect → Recommend) before offering a recommendation. The core insight: the skill's value is not doing more, but doing less — it's a deceleration tool in a landscape of acceleration tools.

Architecture

The entire skill is 215 lines of Markdown in SKILL.md — no executable code, no external dependencies, no state management. It's a pure prompt template deployed as a Claude Code skill.

YAML frontmatter defines the skill name (decision-framework) and a one-line description. The body is an instruction set for Claude organized into four sequential phases:

Phase	Purpose	Lines
1. Understand	What decision, why now, who's affected, desired outcome	45-55
2. Explore	12 topic areas of targeted questions (Purpose through Learning)	58-161
3. Reflect	Neutral summary: assumptions, risks, missing information, strongest arguments	164-174
4. Recommend	Direction + assumptions + confidence level + smallest next step	178-198

Phase gating is explicit: "Only after understanding the context continue" (line 54) and "Only after completing reflection provide" recommendations (line 180). This creates sequential dependency that prevents the model from jumping to solutions prematurely.

The 12 topic areas in Phase 2 provide a taxonomy of question types derived from the 37signals framework: Purpose, Ownership, Time Perspective, Alternatives, Reversibility, Intuition vs Evidence, Consequences, Missing Information, Stakeholders, Values, Effort, Learning. Each area has 3-5 specific prompts — a total of roughly 42 prompting angles, though the skill explicitly says to use only the relevant subset.

The Style section (lines 201-215) is remarkably brief: calm, concise, avoid corporate language, prefer questions over lectures, challenge assumptions respectfully, never overwhelm.

Key Techniques

Adaptive Question Selection

The defining technique. The Core Principles (lines 29-42) explicitly reject the checklist approach:

"Do not dump all questions at once. Instead: understand the decision, ask only the most relevant questions, skip irrelevant ones, keep the conversation flowing naturally."

This trusts the model to calibrate question density to decision complexity — three questions for simple decisions, fifteen for complex ones, never all of them. This is the opposite of most conversational agent designs, which default to exhaustiveness.

Structural Separation of Analysis and Recommendation

The skill separates exploration (Phase 2-3) from recommendation (Phase 4) with an explicit neutrality instruction: "Remain neutral" during reflection. This structural split is more reliable than prompting alone — the model can't accidentally slip into recommendation mode during analysis because the phases create a hard boundary.

Domain Knowledge as Prompt Structure

The 12 topic areas aren't generic "coaching questions" — they're derived from a specific practitioner's framework (37signals' internal decision-making guide). This gives the model concrete angles to explore rather than generic "ask good questions." The taxonomy is the prompt's real intellectual property.

Minimal Viable Constraint

The style guide is four lines. It doesn't over-specify tone, persona, or vocabulary — it trusts the base model's conversational ability and only constrains what matters (calm, concise, non-corporate, Socratic). Compare with coaching skills that spend 50+ lines on voice and persona.

Design Decisions

Prompt template, not code. No API calls, no tools, no persistence. The simplest possible implementation delegates everything to the base model. Sufficient for a coaching conversation; inadequate for tracking decisions over time or verifying claims.

Completeness sacrificed for engagement. The explicit choice to never ask all questions trades systematic coverage for conversation quality. A human coach wouldn't work through a checklist either.

Fixed phase order, no branching. Users who want to jump to a recommendation get redirected to context first. This is philosophically correct (you can't recommend before understanding) but inflexible for return users who want to skip.

No external knowledge integration. The skill can't look up data, check precedents, or verify claims. It's about improving thinking quality, not providing information.

Comparison Notes

vs. Load-Bearing Assumptions: Both are Claude Code skills with structured phase gating. Load-Bearing Assumptions is a multi-agent verification workflow (finder → strategist → parallel validators) with human-in-the-loop checkpoints. Decision Framework is a single-agent conversational skill. One verifies assumptions; the other coaches decisions. The shared pattern is that structured phases prevent the most common LLM failure modes (premature action in one case, premature solutions in the other).

vs. Thought Refiner Skill: Both trust model judgment on output density rather than prescribing fixed counts. Thought Refiner (15 lines) is even more minimal — it sharpens vague input into questions. Decision Framework (215 lines) adds structured coaching through a domain-specific taxonomy. Both share the instinct that less specification is more.

vs. Agent Identity: Agent Identity argues that meaningful interaction requires identity and continuity — a persistent stance. Decision Framework operates within a single session with no persistence. This is a fundamentally different model of value: transient quality improvement vs. persistent relationship.

vs. Generic "Act as a Coach" prompts: The difference is the 12-topic taxonomy. Generic coaching prompts say "ask good questions." Decision Framework provides a specific catalogue of question types from a known practitioner's framework — domain knowledge encoded as instruction structure.

vs. A Smart Bear Skills: Jason Cohen's asb-skills package is the same pattern — a practitioner's business framework encoded as a questioning skill — but pointed in the opposite emotional direction. Decision Framework is neutral coaching ("remain neutral", challenge respectfully); Cohen's Rude Q&A is adversarial, asking unfair questions and refusing vague answers to expose the denial the founder can't self-interrogate away. Both share the restraint-as-feature thesis; only Cohen names a mechanism (denial, excuse-making) for why the questions must be hostile.