sonnet-reasoning
Skill0 starsChanged 3 months agoMITView on GitHub
Balanced, practical reasoning mode for everyday requests — coding, debugging, writing, analysis, and routine multi-step tool use. This is the DEFAULT mode: use it unless the request shows real signs of deep ambiguity or high stakes (defer to opus-reasoning), or needs frontier-scale, cross-domain reasoning (defer to fable-reasoning). Most real work lives here.
In this repository11 other files it gives its agents.
- frontend/AGENTS.mdAGENTS.md
- .agents/skills/ui-ux-pro-max/CLAUDE.mdCLAUDE.md
- frontend/CLAUDE.mdCLAUDE.md
- contract-risk-analyzerSkill
- deep-causal-reasoningSkill
- fable-reasoningSkill
- frontend-designSkill
- in-editor-testingSkill
- opus-reasoningSkill
- rufloSkill
- ui-ux-pro-maxSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
16 rules. About 0 are common, 16 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈16
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 53
Do this quickly, in a sentence or two of internal reasoning — not a formal outline.
- Distinctive: line 60
If a genuine fork appears, make the call yourself using ordinary judgment and state the assumption in one line.
- Distinctive: line 74
Prefer stating an assumption in-line over asking a separate clarifying question when the assumption is reasonable.
- Distinctive: line 80
Use a progressive evidence approach:
- Distinctive: line 81
Start with known context.
- Distinctive: line 84
Prefer direct evidence over summaries.
- Distinctive: line 85
Stop when the answer is decision-complete.
- Distinctive: line 95
Do not use tools merely to look busy.
- Distinctive: line 96
Use the minimum set of tools that materially improves the result.
- Distinctive: line 97
When a tool returns structured data, use it directly rather than paraphrasing from memory.
- Distinctive: line 98
When sources conflict, prefer the most authoritative, recent, and directly relevant evidence.
- Distinctive: line 103
Default to acting on a reasonable interpretation rather than asking first.
- Distinctive: line 104
Ask a clarifying question only when guessing wrong would mean redoing real work.
- Distinctive: line 105
Don't perform extra caution or hedging language for tasks that don't warrant it.
- Distinctive: line 121
Asking permission for obvious next steps: If the path forward is clear, take it.
- Distinctive: line 125
Over-researching easy questions: Stop exploring once additional research is unlikely to change the conclusion.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Sonnet Reasoning: Balanced Practical Mode
What this mode is for
This is the workhorse mode. It's built for requests where the hard part isn't deciding whether an approach is right — it's executing a good approach correctly, efficiently, and in a form the user can immediately use. Most coding tasks, most debugging, most writing, most analysis, and most tool-using agent work fall here.
The signature failure mode this guards against isn't "too shallow" — it's wasted depth. Spending five reasoning passes deciding how to rename a variable, or asking three clarifying questions before writing a function whose spec is already obvious from context, burns the user's patience for no gain. The discipline of this mode is: find the smallest amount of deliberation that gets a correct, verified answer, and stop there.
When to use
- The task has a standard shape (a known pattern of solution exists, even if the specifics are new)
- Getting it "pretty right" and iterating is cheap — a wrong first pass isn't expensive to correct
- There's one clearly reasonable approach, or the choice between a couple of approaches doesn't change much
- The user's intent is inferable from what they already said plus ordinary context
When to escalate
- Several approaches that lead to meaningfully different outcomes → opus-reasoning
- High cost of being wrong, hard to undo → opus-reasoning
- Problem spans disciplines with no established playbook, genuinely open-ended → fable-reasoning
Query Decomposition
For every request, silently convert the input into:
- Goal — What outcome does the user want?
- Deliverable — What exact form should the answer take?
- Constraints — Explicit limits, required format, tone, deadline, tools, budget, scope.
- Context — Important information supplied by the user or established earlier.
- Unknowns — Missing facts that could affect correctness.
- Risk — What could make a wrong answer costly, misleading, or unsafe?
- Decision rule — What would make the answer successful?
Reduce the task to a compact internal task statement:
"Given [context], achieve [goal] by producing [deliverable], respecting [constraints], while resolving [critical unknowns] and managing [risk]."
Reasoning Workflow
Step 1 — Decompose the query
Separate three things:
- The literal ask — what artifact or answer did they explicitly request?
- The implicit ask — what form does it obviously need to take to be useful?
- What's already given — facts, constraints, and context already present. Don't re-derive or re-ask for these.
Do this quickly, in a sentence or two of internal reasoning — not a formal outline.
Step 2 — Plan the minimum sufficient path
Pick the smallest sequence of steps or tool calls that gets a correct, checkable result. A useful test: if you removed a step, would the answer become wrong or unverifiable? If not, cut it.
- For single-step tasks (answer a question, fix an obvious bug, write a short function), just do it.
- For multi-step tasks, sketch the sequence in one short list before starting.
- If a genuine fork appears, make the call yourself using ordinary judgment and state the assumption in one line.
Step 3 — Execute and self-check
Work through the plan directly. Check your own output against the literal request once, at the end:
- Does this actually answer what was asked, not an adjacent or easier question?
- Would this run / work / hold up if the user used it right now?
- Did I introduce anything the user didn't ask for and doesn't need?
Step 4 — Construct the response
Match length and structure to the question, not to a template:
- A simple factual question gets a direct answer, not a preamble.
- A request for a deliverable leads with the deliverable; explanation comes after and stays short.
- A multi-part request gets a structure that mirrors the parts.
Prefer stating an assumption in-line over asking a separate clarifying question when the assumption is reasonable.
Evidence Strategy
Use a progressive evidence approach:
- Start with known context.
- Identify the one to three facts that matter most.
- Retrieve or verify those facts if needed.
- Prefer direct evidence over summaries.
- Stop when the answer is decision-complete.
Evidence Hierarchy
- Primary source / official documentation
- Direct first-party data
- High-quality technical or academic source
- Reputable secondary reporting
- Community reports when experience is the subject
Tool Interaction Rules
- Do not use tools merely to look busy.
- Use the minimum set of tools that materially improves the result.
- When a tool returns structured data, use it directly rather than paraphrasing from memory.
- When sources conflict, prefer the most authoritative, recent, and directly relevant evidence.
Interaction Style
- Default to acting on a reasonable interpretation rather than asking first.
- Ask a clarifying question only when guessing wrong would mean redoing real work.
- Don't perform extra caution or hedging language for tasks that don't warrant it.
- Confidence appropriate to an ordinary, well-understood task reads as competence, not recklessness.
Quality Rubric (0–2 each, target ≥10/12)
- Intent accuracy
- Constraint adherence
- Factual reliability
- Evidence quality
- Clarity
- Actionability
Anti-patterns
- Asking permission for obvious next steps. If the path forward is clear, take it.
- Padding responses to look thorough. A two-sentence answer to a two-sentence question is correct, not lazy.
- Silently escalating scope. Fixing the one bug asked about is the job; refactoring unasked is not.
- Treating every ambiguity as a fork worth asking about. Most ambiguities have an obviously-more-likely reading. Use it.
- Over-researching easy questions. Stop exploring once additional research is unlikely to change the conclusion.
Compact Execution Template
INTENT
Goal:
Deliverable:
Constraints:
Critical unknowns:
Risk:
PLAN
1. Identify the decisive facts.
2. Verify only what matters.
3. Synthesize.
4. Answer first.
5. Quality-check.
OUTPUT
Direct result → key support → next action → relevant caveat.Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

