thought-based-reasoning
À propos
Cette compétence de Claude permet un raisonnement explicite et étape par étape pour des tâches complexes telles que les mathématiques à plusieurs étapes, les énigmes logiques et l'analyse des compromis. Elle s'active face à une analyse superficielle, des erreurs arithmétiques ou des incertitudes, forçant le modèle à décomposer les problèmes pour améliorer la précision. Les développeurs doivent l'utiliser pour montrer un travail vérifiable et détecter les erreurs plutôt que de sauter directement aux réponses.
Installation rapide
Claude Code
Recommandénpx skills add guia-matthieu/clawfu-skills -a claude-code/plugin add https://github.com/guia-matthieu/clawfu-skillsgit clone https://github.com/guia-matthieu/clawfu-skills.git ~/.claude/skills/thought-based-reasoningCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Thought-Based Reasoning
Overview
Core principle: Making reasoning explicit improves accuracy 20-70% on complex tasks.
Instead of jumping to answers, decompose problems into steps. This catches errors, enables backtracking, and produces verifiable reasoning.
When to Use
digraph decide {
"Problem type?" [shape=diamond];
"Direct answer worked?" [shape=diamond];
"Need confidence?" [shape=diamond];
"Use direct prompting" [shape=box];
"Use Zero-shot CoT" [shape=box];
"Use Self-Consistency" [shape=box];
"Use technique from table" [shape=box];
"Problem type?" -> "Direct answer worked?" [label="simple"];
"Problem type?" -> "Use technique from table" [label="math/logic/creative"];
"Direct answer worked?" -> "Use direct prompting" [label="yes"];
"Direct answer worked?" -> "Need confidence?" [label="no"];
"Need confidence?" -> "Use Self-Consistency" [label="yes, high stakes"];
"Need confidence?" -> "Use Zero-shot CoT" [label="no, just need better"];
}
Use when:
- Multi-step arithmetic or word problems
- Logic requiring deduction chains
- Decisions with multiple factors
- Creative problems needing exploration
- Any task where direct answer was wrong
Don't use when:
- Simple factual recall
- Single-step operations
- Time-critical responses where accuracy tradeoff acceptable
Quick Reference
| Technique | Trigger | Template |
|---|---|---|
| Zero-shot CoT | Quick reasoning boost | "Let's think step by step..." |
| Self-Consistency | High-stakes decision | Run 3-5 paths, majority vote |
| Tree of Thoughts | Puzzle/creative block | Branch, evaluate, backtrack |
| Least-to-Most | Complex multi-part problem | Decompose → solve subproblems → combine |
| ReAct | Need external facts | Thought → Action → Observation loop |
| PAL | Math with computation | Generate code, execute it |
Techniques
1. Zero-shot Chain-of-Thought
When: Quick prototype, no examples available
Template:
[Problem statement]
Let's think step by step:
Example:
A store has 45 apples. They sell 12 in the morning and receive a shipment of 30.
Then they sell 18 more. How many apples remain?
Let's think step by step:
1. Start: 45 apples
2. Sell 12: 45 - 12 = 33 apples
3. Receive 30: 33 + 30 = 63 apples
4. Sell 18: 63 - 18 = 45 apples
Answer: 45 apples remain.
Accuracy gain: +20-60%
2. Self-Consistency
When: High-stakes decisions, need confidence measure
Process:
- Run Zero-shot CoT 3-5 times (vary temperature if possible)
- Collect all final answers
- Take majority vote
- Report confidence as agreement ratio
Template:
[Problem]
I'll reason through this multiple ways to verify:
Path 1:
[reasoning...]
Answer: X
Path 2:
[reasoning...]
Answer: Y
Path 3:
[reasoning...]
Answer: X
Consensus: X (2/3 agreement = 67% confidence)
Accuracy gain: +10-20% over single CoT
3. Tree of Thoughts
When: Puzzles, creative problems, need to explore alternatives
Process:
- Generate 2-3 initial approaches
- Evaluate each (promising/uncertain/dead-end)
- Expand promising branches
- Backtrack from dead-ends
- Continue until solution found
Template:
[Problem]
## Branch 1: [Approach A]
Evaluation: [promising/uncertain/dead-end]
[If promising, continue...]
## Branch 2: [Approach B]
Evaluation: [promising/uncertain/dead-end]
[If dead-end, note why and stop]
## Expanding Branch 1:
### Branch 1.1: [Sub-approach]
...
## Solution found in Branch 1.1
Example (Game of 24: make 24 from 4, 7, 8, 8):
Branch 1: Try multiplication first
- 4 × 7 = 28... need to subtract 4, but only have 8,8
- Evaluation: uncertain, continue
Branch 2: Try getting 3 × 8 = 24
- Need to make 3 from 4, 7, 8
- 7 - 4 = 3 ✓
- 3 × 8 = 24 ✓
- Evaluation: SOLUTION
Answer: (7 - 4) × 8 = 24 (using one 8, ignoring other)
Wait - must use all numbers. Backtrack.
Branch 3: Use both 8s
- 8 + 8 = 16, need +8 from 4,7... 7-4=3, not 8
- 8 × 8 = 64, need to subtract 40 from 4,7... no
- 8 - 8 = 0, useless
- 8 / 8 = 1, then 4 × 7 × 1 = 28, not 24
Branch 4: Division approach
- (8 - 4) = 4
- 8 / 4 = 2... not useful
- 4 / (8-8) = undefined
Branch 5: Fractions
- 8 / (1 - 7/8) = 8 / (1/8) = 64... no
- Actually: 8 × (7 - 8/4) = 8 × (7-2) = 8 × 5 = 40
Hmm, trying more systematically...
- Need operations on 4,7,8,8 = 24
- (4 - 8/8) × 7 = 3 × 7 = 21
- 4 × (8 - 8/7)... messy
- (8 - 4) × (8 - 2)... don't have 2
Let me try: 8 / (8 - 4/7)... no
Try: (7 + 8/8) × 4 = 8 × 4 = 32... nope that's wrong
Wait: (7 + 1) × 3 = 24, can I make 1 and 3?
8/8 = 1 ✓, but can't make 3 from just 4
Try factoring 24: 1×24, 2×12, 3×8, 4×6
- 3×8: need 3 from 4,7,8. (4+8)/7 ≈ 1.7, 7-4=3 ✓ but used 8 for the ×8
- 4×6: need 6 from 7,8,8. 8-8+7=7, (8+8)/7≈2.3, nope
Solution: (8 - 8/4) × 7 = (8 - 2) × 7 = 6 × 7 = 42... no
Actually: 4 × 8 - 8 + 7 - 7 = 32... wait no sevens wrong
Let me be systematic: 4 × 8 = 32, 32 - 8 = 24 ✓
What about 7? 32 - 8 × 7/7 = 32 - 8 = 24 ✓
Answer: 4 × 8 - 8 × 7/7 = 32 - 8 = 24
Or simpler: 4 × 8 - 8 + 7 - 7 = 24 (trivially using 7-7=0)
Accuracy gain: +50-70% on hard puzzles
4. Least-to-Most Prompting
When: Complex problem with subproblems
Process:
- Decompose into subproblems
- Solve easiest first
- Use solutions to solve harder ones
- Combine for final answer
Template:
[Complex problem]
## Subproblems (easiest to hardest):
1. [Subproblem A]
2. [Subproblem B, may need A's answer]
3. [Subproblem C, needs A and B]
## Solutions:
### Subproblem 1:
[solve...]
Answer: [X]
### Subproblem 2 (using X):
[solve...]
Answer: [Y]
### Subproblem 3 (using X, Y):
[solve...]
## Final Answer:
[Combine solutions]
Accuracy gain: +30-80% on compositional tasks
5. ReAct (Reasoning + Acting)
When: Need external information, reduce hallucination
Process:
- Thought: reason about what's needed
- Action: query external source
- Observation: record result
- Repeat until solved
Template:
Question: [Question requiring external info]
Thought 1: I need to find [X] to answer this.
Action 1: Search/Lookup [X]
Observation 1: [Result]
Thought 2: Now I know X. I also need [Y].
Action 2: Search/Lookup [Y]
Observation 2: [Result]
Thought 3: With X and Y, I can now answer.
Answer: [Final answer grounded in observations]
Accuracy gain: +15-35%, major hallucination reduction
6. PAL (Program-Aided Language)
When: Math with computation, eliminate arithmetic errors
Process:
- Translate problem to code
- Execute code
- Return result
Template:
[Math problem]
Let me write code to solve this:
```python
# [Problem restated as comments]
initial = 45
after_morning_sales = initial - 12
after_shipment = after_morning_sales + 30
after_afternoon_sales = after_shipment - 18
print(f"Remaining: {after_afternoon_sales}")
[Execute] Output: Remaining: 45
Answer: 45
**Accuracy gain:** Eliminates arithmetic errors entirely
## Decision Matrix
| Situation | Best Technique |
|-----------|----------------|
| Quick reasoning, no examples | Zero-shot CoT |
| High-stakes, need confidence | Self-Consistency |
| Puzzle, creative, exploration needed | Tree of Thoughts |
| Multi-part with dependencies | Least-to-Most |
| Need facts, reduce hallucination | ReAct |
| Math with many calculations | PAL |
| Iterative improvement | Reflexion (run, critique, retry) |
## Common Mistakes
| Mistake | Fix |
|---------|-----|
| Using CoT for simple queries | Direct answer is fine for 1-step problems |
| Not showing work | Explicit steps catch errors |
| Stopping at first answer | Self-consistency finds better answers |
| Linear thinking on puzzles | Tree of Thoughts enables backtracking |
| Computing mentally | PAL eliminates arithmetic errors |
| Guessing facts | ReAct grounds in external sources |
## Combining Techniques
For maximum accuracy on hard problems:
- Least-to-Most: decompose into subproblems
- For each subproblem:
- PAL if computational
- ReAct if needs facts
- Tree of Thoughts if exploratory
- Self-Consistency on final assembly
---
## What Claude Does vs What You Decide
| Claude handles | You provide |
|---------------|-------------|
| Selecting appropriate reasoning technique | Problem statement and constraints |
| Executing multi-step reasoning chains | Verification of intermediate steps |
| Generating multiple reasoning paths | Selection of best answer |
| Backtracking from dead-ends | Judgment on acceptable confidence |
| Computing via PAL when needed | Real-world validation of results |
---
## Skill Boundaries
### This skill excels for:
- Math and logic problems with multiple steps
- Decisions with competing factors
- Puzzles requiring exploration
- Tasks where initial answers were wrong
### This skill is NOT ideal for:
- Simple factual recall → Direct answer is faster
- Creative writing → Different techniques apply
- Time-critical responses → CoT adds latency
---
## Skill Metadata
```yaml
name: thought-based-reasoning
category: thinking
version: 2.0
author: GUIA
source_expert: Wei et al. (CoT), Yao et al. (ToT), Kojima et al. (Zero-shot CoT)
difficulty: intermediate
mode: both
tags: [reasoning, cot, tot, react, pal, logic, math, problem-solving]
created: 2026-02-03
updated: 2026-02-03
Dépôt GitHub
Compétences associées
llamaguard
AutreLlamaGuard est le modèle de Meta, doté de 7 à 8 milliards de paramètres, conçu pour modérer les entrées et sorties des LLM selon six catégories de sécurité comme la violence et les discours haineux. Il offre une précision de 94 à 95 % et peut être déployé avec vLLM, Hugging Face ou Amazon SageMaker. Utilisez cette compétence pour intégrer facilement le filtrage de contenu et des garde-fous de sécurité dans vos applications d'IA.
cost-optimization
AutreCette compétence de Claude aide les développeurs à optimiser les coûts du cloud grâce au redimensionnement des ressources, aux stratégies d'étiquetage et à l'analyse des dépenses. Elle fournit un cadre pour réduire les dépenses cloud et mettre en œuvre une gouvernance des coûts sur AWS, Azure et GCP. Utilisez-la lorsque vous devez analyser les coûts d'infrastructure, redimensionner les ressources ou respecter des contraintes budgétaires.
quantizing-models-bitsandbytes
AutreCette compétence quantifie les LLMs en précision 8 bits ou 4 bits à l'aide de bitsandbytes, permettant une réduction de 50 à 75 % de la mémoire utilisée avec une perte de précision minime. Elle est idéale pour exécuter des modèles plus volumineux sur une mémoire GPU limitée ou pour accélérer l'inférence, prenant en charge des formats comme INT8, NF4 et FP4. La compétence s'intègre à HuggingFace Transformers et permet l'entraînement QLoRA ainsi que l'utilisation d'optimiseurs en 8 bits.
dispatching-parallel-agents
AutreCette compétence Claude déploie plusieurs agents pour enquêter et résoudre simultanément 3 problèmes indépendants ou plus. Elle est conçue pour des scénarios impliquant des défaillances non liées qui peuvent être résolues sans état partagé ni dépendances. La capacité fondamentale est la résolution de problèmes en parallèle, en assignant un agent par domaine problématique indépendant afin de maximiser l'efficacité.
