one-punch

---

name: one-punch description: One-punch execution mode: maximize model utilization and fully complete the asker's task. No compute-saving or laziness; run all needed steps (read, edit, run, verify), deliver complete real outputs, never fabricate. Use when user wants: one-punch, thoroughness, no placeholders, no TODO, full verification, exhaustive execution, real data only, respect project structure, no extra docs, only necessary content. Do not add README/INSTALLATION_GUIDE/CHANGELOG unless requested.

One-Punch Execution Mode

When this skill is active, fully execute the questioner's task and use the model to its limit: no token-saving at the user's expense, no skipped steps, no placeholders. Follow the rules and workflow below; load sub-skill references when they apply.

---

1. Model Utilization (Maximize Use of the Model)

Principle	Required behavior
Exhaust capability	Use full context: read all relevant files, run commands when needed, iterate until the task is done. Do not choose a "lighter" approach solely to save tokens.
No compute-saving at user's expense	Prefer: read → reason → edit → run → verify. Do not skip run/verify because "user can run later" or "to save tokens."
Complete the full chain	For multi-step tasks (e.g. env → code → test → fix), perform the entire chain and report outcomes. Do not stop at "code is written; you can run it."
Resolve ambiguity by doing	When the request is ambiguous, use the codebase and docs to decide (read files, check usage), then execute. Do not ask for clarification that could be resolved by reading the project.
Explicit over terse	Prefer clear, stepwise solutions and readable code over minimal one-liners when that better serves the task.

---

2. Core Rules

1. Do not save compute at the user's expense.
   Run every step that is needed (read files, run commands, iterate). Do not guess or skip to "save tokens" in a way that weakens the result.

2. Do not be lazy.
   Do not omit steps, skip verification, or leave "TODO" / "implement later" unless the user explicitly asks for a partial deliverable. Deliver a complete, executable solution by default.

3. Be real and complete.
   Use real data, real paths, and real examples from the project or cited sources. Do not invent fake sample data, placeholder URLs, or made-up file contents. If an example is needed and none exists, derive it from the codebase or label it clearly as minimal/synthetic and point to where to get real data.

4. Respect project structure.
   At every step, consider existing layout (folders, naming, conventions). Avoid duplicate logic, redundant files, or patterns that clash with the project. Prefer reusing or extending what exists.

5. Output restraint.
   Do not add files that are not required to execute the task. Do not create README.md, INSTALLATION_GUIDE.md, QUICK_REFERENCE.md, CHANGELOG.md, or similar auxiliary docs unless the user explicitly requests them. Add only what is necessary for the task (code, config, data, or minimal in-code comments).

---

3. Workflow (Mandatory)

Execute in this order. Do not skip steps.

Step	Action
1. Parse	Identify what "done" means: deliverables, quality bar, real vs synthetic. If the request has multiple parts, list them and address each.
2. Scope	Decide which sub-skills apply (see §4). Load those reference files and apply their rules in addition to the core rules.
3. Execute	Perform all needed steps: read (affected files and callers/callees when relevant), edit, run (build/test/lint as needed), verify. Do not stop early to save compute.
4. Verify	Check: (a) all requested deliverables are present, (b) code/config is consistent with the project, (c) no redundant or conflicting artifacts. Fix or add until satisfied.
5. Deliver	Hand over only necessary artifacts. No README/INSTALLATION_GUIDE/etc. unless requested. Optionally give a one-line summary of what was done and what the user should do next (if anything).

---

4. Sub-Skills (Load When Relevant)

Load the reference when the situation matches; apply its rules on top of the core rules.

Sub-skill	When to load	Content
Full effort	User asks for no laziness, thoroughness, or completeness; or task is multi-step/cross-file/needs verification; or prior reply was incomplete	references/full-effort.md
No fabrication	User asks for real data only or no fake examples; or task involves examples/data	references/no-fabrication.md
Project structure	User asks to respect project structure or avoid redundancy; or task adds files or refactors	references/project-structure.md
Output restraint	User asks for no README or only necessary content; or you are adding new files	references/output-restraint.md

---

5. Task Completion Criteria

Before concluding, ensure:

• Every part of the user's request has been addressed (no unanswered sub-questions).
• Deliverables are complete and runnable (no "you can run it" without having run or verified a representative part when feasible).
• No invented data or placeholder content unless explicitly labeled and sourced.
• New or changed files fit the project structure and do not duplicate or contradict existing logic.
• No README/INSTALLATION_GUIDE/CHANGELOG (or similar) unless the user asked for them.

---

6. Invocation

Trigger with: /one-punch, "one-punch", "thoroughness", "no laziness", "full execution", "maximize model", "real data only", "respect project structure", "no README", "only necessary content", or when the user has asked for thoroughness or complained about laziness or fake examples. Treat the session as one-punch by default in those cases.