AGENT SKILL

guide-decision

Use when the user asks for personal, non-technical advice with no single correct answer - personal finance, veterinary or pet care choices, management or interpersonal decisions, legal-adjacent personal questions, or other life decisions. Also covers method and delivery judgment calls (estimation approach, planning cadence, process tradeoffs). Not for technical/architecture work, or questions with one objectively correct answer. Personal purchases of software/subscriptions route to `guide-purchase` instead.

License 许可
BSD-3-Clause
最近更新
2026-11-11
Stars
★ 1
Forks
0

01 前言 Front Matter

name
guide-decision
version
1.1.0
description
Use when the user asks for personal, non-technical advice with no single correct answer - personal finance, veterinary or pet care choices, management or interpersonal decisions, legal-adjacent personal questions, or other life decisions. Also covers method and delivery judgment calls (estimation approach, planning cadence, process tradeoffs). Not for technical/architecture work, or questions with one objectively correct answer. Personal purchases of software/subscriptions route to `guide-purchase` instead.

03 SKILL.md 正文

Guide Decision

Overview

An unframed question has no right answer, and a confident first answer is a draft, not a verdict. Framing before the answer, then a deliberate self-review, catches the assumptions and alternatives a single pass misses.

Core principle: Never answer a question you have not finished framing - then never ship the first plausible answer either.

When to Use

  • The user asks "what should I do about X" for a personal, non-technical matter
  • The user asks for an opinion or recommendation where reasonable experts could disagree
  • The question has real tradeoffs (cost vs. time, risk vs. comfort, short vs. long term)

Not for:

  • Technical/architecture work - how to build, structure, or fix a system. Method and delivery calls (how to estimate, how to plan, which process to run) stay in scope, as does a personal software or subscription purchase, via guide-purchase
  • Questions with one factually correct answer (use verify-sources instead)
  • Buying decisions (use guide-purchase - it reuses the loop below)

Core Pattern

  1. Frame the need - a grilling gate, run before answering: every ambiguity, unknown, or assumption whose answer would change the recommendation becomes a question, batched by independent branch with a recommended answer each. Concrete is not specified - "how do I service my Peugeot 207" is a perfectly clear ask that still needs year, engine and trim before any answer, or any verification behind it, holds. Ask for the identifier or the constraint - model reference, existing setup, jurisdiction, horizon - never a photo or a full inventory. Skip only what genuinely cannot move the answer, and don't re-ask what a same-domain followup already established this session.
  2. Answer as a named domain expert - state the persona explicitly ("As a [domain] expert...") and give a direct, complete first-pass answer. Define every domain term on first use, and carry one concrete worked example through the answer: a term used without definition, or a formula without numbers, is not an expert answer. On a same-domain followup in the same session, don't reintroduce the persona - answer directly using the context already established.
  3. Pause before finalizing - take the time to reflect. Don't finalize the first plausible answer; check it actually addresses what was asked.
  4. Self-critique the draft - list the assumptions made, and surface 1-3 credible alternatives or considerations the first pass missed. Scale the depth to the stakes: a quick personal question needs 1-2 lines, not a multi-section audit; a same-domain followup needs this only if it adds a new tradeoff. When the stakes involve legal exposure, health, or losing something hard to reverse (housing, a job), also flag any specific fact/figure/threshold stated with confidence in step 2 that isn't verified - point to verify-sources or a relevant professional rather than asserting it.
  5. Revise - rewrite the answer incorporating the critique. Merge it into a stronger answer; don't tack the critique on as caveats. If the reasoning surfaced a variable you guessed that step 1 could not have known to ask about, ask it now and mark the recommendation provisional on it - not a second round of framing.
  6. Close with what would change this - one line naming the condition(s) that would flip the recommendation, so it stays revisable as facts change.
  7. Consolidate - when the frame or the recommendation has materially changed across refinement rounds, re-issue the full answer as one document instead of another increment. Deltas accumulate into an unusable answer.

If the user asks to skip the ceremony ("just a direct answer", "no analysis") - steps 3-7 still happen, but stay internal: reply with only the revised answer, dropping the visible critique and closing line. Step 1 is not covered by this: an unasked question that changes the answer is not ceremony. Don't skip the critique itself, only its visible form.

Quick Reference

Step Purpose Skip when
Frame (grilling) Close every ambiguity/unknown/assumption that moves the answer Nothing left that moves the answer, or already established this session
Expert answer Anchor a direct first pass; terms defined, example worked Persona intro skipped on a same-domain followup (the answer itself never is)
Reflect Interrupt rushing to the first answer Never
Self-critique Surface assumptions + alternatives; verify facts if high-stakes Never*
Revise Merge critique into final answer; ask what only the reasoning revealed Never
What would change it Keep the recommendation revisable Never*
Consolidate Re-issue in full once the frame or recommendation materially moved The increment still reads on its own

*Visible form only - stays internal if the user asked for a direct answer.

Example

User: "Should I put my dog on a raw diet?"

Frame first - three answers change the recommendation, so they get asked before it is written, not assumed:

  • Age and health status? (Recommended answer: healthy adult. Puppies, seniors and immunocompromised dogs flip the risk calculus.)
  • Has a vet already recommended raw for a specific condition? (Recommended: no - if yes, their guidance supersedes this.)
  • Anyone immunocompromised, pregnant, or under 5 in the household? (Recommended: no - it drives the handling risk, not the dog's.)

User: healthy 4-year-old, no vet recommendation, two adults at home.

As a veterinary nutrition expert: raw diets carry salmonella/listeria risk for the pet and household, and are rarely nutritionally complete without careful formulation - I'd recommend a commercial diet formulated to AAFCO standards instead, with real meat high on the ingredient list.

Alternative not yet covered: a vet-supervised, commercially-balanced raw diet avoids most of the nutritional-completeness risk, if raw is important to the owner.

Revised: for a healthy dog, a quality AAFCO-formulated commercial diet is the lower-risk default. If raw still matters to you, do it only with a vet-formulated recipe (not homemade improvisation) to avoid the deficiency risk that makes raw diets risky in the first place.

This would change if a vet has already diagnosed a condition raw feeding is meant to address - follow their specific guidance instead.

Common Mistakes

  • Treating a concrete ask as a specified one - skipping the frame because the question reads clearly, when the answer still pivots on an unstated model, year, variant, or existing setup
  • Answering about a specific instance from a generic mental model when variants differ - describing a mechanism, or costing a setup, the user's actual model may not have
  • Stacking delta after delta across refinement rounds without ever re-issuing a consolidated answer
  • Self-critiquing but not actually revising - the critique becomes a hedge appended to the original answer instead of changing it
  • Naming a persona without answering as that persona - "As a vet, it depends" isn't an expert answer
  • Producing alternatives that are cosmetic variations of the same answer, not genuinely different considerations
  • Writing a multi-section visible audit trail for a low-stakes question - match the critique's length to how much the decision matters
  • Treating "just a direct answer" as license to skip the internal critique itself, not just its visible form
  • Re-framing or reintroducing the persona on a same-domain followup already established earlier in the session
  • Stating a specific legal threshold, dosage, or figure with confidence on a high-stakes question without flagging it for verification

04 同仓库更多 Skills

NormandErwan
NormandErwan/Claude

write-guide

Use when writing a guide someone else will follow - a method guide, a playbook, a how-we-do-it document for a team or a client - or when reworking, auditing or extending one that already exists. Also when an outline is handed to you to fill in, when a guide's rules need evidence labels, or when a second guide joins the first and the two start contradicting each other.

★ 1
NormandErwan
NormandErwan/Claude

write-french

Use when the output is French - a chat reply in French, or a French markdown deliverable such as a guide, a report or a synthesis - including when the user writes in French and expects French back. Prevents French that reads as translated English: word-for-word calques, sentences compressed until the meaning is gone, jargon used before it is defined, invented turns of phrase.

★ 1
NormandErwan
NormandErwan/Claude

verify-sources

Use when about to state a factual, technical, historical, or procedural claim in ANY domain (software, DIY/craft, health, legal, finance...) - a version number, default value, statistic, API behavior, a legal deadline, a health claim, a manufacturer spec, "X always/never does Y", "the docs say", "best practice is" - that you have not verified against a source this turn, or when tempted to answer a factual question from memory under time pressure

★ 1
NormandErwan
NormandErwan/Claude

token-file-ops

Use when reading, editing, creating, or manipulating files; or running commands with potentially large output. Provides generic bash-first patterns for any file type.

★ 1
NormandErwan
NormandErwan/Claude

token-codebase-exploration

Use when exploring an unfamiliar codebase, understanding how a specific feature or pattern works, or learning how components interact. Provides the exploration workflow and strategy.

★ 1
NormandErwan
NormandErwan/Claude

teach

Use when the user sets learning a topic or skill as their own goal - "teach me X", "I want to learn X", "help me get good at X", a course or study plan for themselves. Not for a question one reply answers ("what is X", "explain X").

★ 1