Przejdź do treści
Biblioteka skilli
AI i agenci · Pisanie

prompt-engineering

/prompt-engineering

Pomaga tworzyć, poprawiać i debugować prompty, łańcuchy promptów i prompty systemowe oraz dostosowywać je do GPT, GLM i Gemini albo migrować między wersjami modeli, w tym z Claude 4 do Claude 5.

Repozytorium
ai-tools ↗
Licencja
MIT
Ostatnia zmiana repozytorium
26.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
43
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Pomaga tworzyć, poprawiać i debugować prompty, łańcuchy promptów i prompty systemowe oraz dostosowywać je do GPT, GLM i Gemini albo migrować między wersjami modeli, w tym z Claude 4 do Claude 5.

Oryginalny opis (EN)

Use when the user asks to create, write, improve, or debug a prompt, build a prompt chain, design a system prompt, adapt a prompt for GPT-6 or GPT-5.6 (OpenAI), GLM 4.7 (Z.ai), or Gemini 3, create a Gemini deep research prompt, migrate a prompt or skill from Claude 4 to Claude 5, from Opus 5 to Opus 5.5, or between GPT versions (e.g. GPT-5.6 to GPT-6), optimize LLM-targeted content (skills, agents, instructions, documentation), author a prompt an invoking workflow dispatches to a worker process, distill prompt-authoring rules into a session-scoped ruleset, or needs prompt engineering guidance.

Jak działa

  1. Prompt Recognition
  2. Workflow
  3. Essential Techniques Reference
  4. Model-Generation Optimizations
  5. Claude 4 → Claude 5 Migration (When Requested)
  6. GPT-6 / GPT-5.6 (OpenAI) Adaptation (When Requested)
  7. GLM 4.7 Adaptation (When Requested)
  8. Gemini 3 Adaptation (When Requested)
  9. Output Format
  10. Quality Checklist
Sekcje z nagłówków SKILL.md, w oryginale · 383 linii, 3 438 słów
Podgląd SKILL.md
---
name: prompt-engineering
version: 3.13.0
description: Use when the user asks to create, write, improve, or debug a prompt, build a prompt chain, design a system prompt, adapt a prompt for GPT-6 or GPT-5.6 (OpenAI), GLM 4.7 (Z.ai), or Gemini 3, create a Gemini deep research prompt, migrate a prompt or skill from Claude 4 to Claude 5, from Opus 5 to Opus 5.5, or between GPT versions (e.g. GPT-5.6 to GPT-6), optimize LLM-targeted content (skills, agents, instructions, documentation), author a prompt an invoking workflow dispatches to a worker process, distill prompt-authoring rules into a session-scoped ruleset, or needs prompt engineering guidance. Also handles prompt refinement and follow-up modifications.
---

# Prompt Engineering Lab

**Deliverable**: Deliver the artifact and never carry out the task it describes — a prompt, or the ruleset an invoking session authors prompts from. Its recipient is a person who copies it, a process an invoking workflow passes it to, or the invoking session itself; dispatching it is that workflow's own next step, not execution by this skill. Shape the delivery to the recipient via the [Output Format](#output-format) table.

## Prompt Recognition

A prompt is any content containing information, context, or directives intended for LLM consumption. This includes:

**Traditional prompts:**
- System prompts, user prompts, prompt templates
- Few-shot examples, prompt chains

**LLM-targeted content:**
- Skills, agents, commands (Claude Code, Cursor, etc.)
- Project instructions and rules files
- Documentation consumed by LLMs during tasks
- Any file with imperative instructions for an LLM

**Recognition signals:**
- Frontmatter with fields like `description`, `tools`, `allowed-tools`
- Imperative language: "You must", "Always", "Never"
- Workflow steps, constraints, decision trees
- References to LLM tools or capabilities
- Content structured to guide LLM behavior

**Optimization principle for LLM-targeted content:**

These documents are created with clear scope and intention—often by an LLM—and contain everything required for their purpose. Optimization improves clarity, structure, and effectiveness WITHOUT removing information, context, or directives.

Apply reasoning to understand:
- What information is essential to the document's purpose
- Which directives shape LLM behavior
- What context enables correct interpretation

Then optimize for clarity and impact while preserving all substantive content. Personas and "You are …" identity openers are not substantive content.

Because Claude 5 is the default target, apply "less is more" by default: Claude 5 exercises judgment, so cutting over-constraint (absolute rules, repetition, worked examples that narrow exploration) improves output. This relaxes the preserve-everything default above — keep substantive directives, but drop scaffolding that compensated for weaker instruction-following. For Claude 4 or earlier targets, keep the preserve-everything default. See `references/claude-5-guide.md#optimizing-llm-targeted-content-for-claude-5`.

## Workflow

```dot
digraph prompt_engineering {
    "Prompt request" [shape=doublecircle];
    "Refining a prompt already generated this conversation?" [shape=diamond];
    "Refinement mode: ask what changes, deliver the full refined prompt" [shape=box];
    "Migrating existing Claude 4 content to Claude 5?" [shape=diamond];
    "Migration mode: re-tune per claude-5-guide, deliver before/after" [shape=box];
    "Distilled ruleset for the invoking session to author prompts from?" [shape=diamond];
    "Ruleset mode: derive rules from the target-model guide and the session's artifact types" [shape=box];
    "Traditional prompt or LLM-targeted content?" [shape=diamond];
    "Phase 1: scope the prompt (purpose, audience, success criteria, platform, target model)" [shape=box];
    "Phase 2: design strategy (techniques by complexity + target-model adaptation)" [shape=box];
    "Target model?" [shape=diamond];
    "Claude 5 default -> claude-5-guide" [shape=box];
    "Claude 4 -> claude-4-guide" [shape=box];
    "GPT-6 / GPT-5.x (OpenAI) -> openai-gpt-guide" [shape=box];
    "GLM 4.7 -> glm-47-guide" [shape=box];
    "Gemini 3 -> gemini-3-guide (Deep Research -> gemini-3-deep-research-guide)" [shape=box];
    "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?" [shape=diamond];
    "Phase 3: deliver the artifact (format by type, recipient, and platform)" [shape=doublecircle];

    "Prompt request" -> "Refining a prompt already generated this conversation?";
    "Refining a prompt already generated this conversation?" -> "Refinement mode: ask what changes, deliver the full refined prompt" [label="yes"];
    "Refinement mode: ask what changes, deliver the full refined prompt" -> "Phase 3: deliver the artifact (format by type, recipient, and platform)";
    "Refining a prompt already generated this conversation?" -> "Migrating existing Claude 4 content to Claude 5?" [label="no"];
    "Migrating existing Claude 4 content to Claude 5?" -> "Migration mode: re-tune per claude-5-guide, deliver before/after" [label="yes"];
    "Migration mode: re-tune per claude-5-guide, deliver before/after" -> "Phase 3: deliver the artifact (format by type, recipient, and platform)";
    "Migrating existing Claude 4 content to Claude 5?" -> "Distilled ruleset for the invoking session to author prompts from?" [label="no"];
    "Distilled ruleset for the invoking session to author prompts from?" -> "Ruleset mode: derive rules from the target-model guide and the session's artifact types" [label="yes"];
    "Ruleset mode: derive rules from the target-model guide and the session's artifact types" -> "Target model?";
    "Distilled ruleset for the invoking session to author prompts from?" -> "Traditional prompt or LLM-targeted content?" [label="no"];
    "Traditional prompt or LLM-targeted content?" -> "Phase 1: scope the prompt (purpose, audience, success criteria, platform, target model)" [label="traditional"];
    "Traditional prompt or LLM-targeted content?" -> "Phase 2: design strategy (techniques by complexity + target-model adaptation)" [label="LLM-targeted (skip Phase 1)"];
    "Phase 1: scope the prompt (purpose, audience, success criteria, platform, target model)" -> "Phase 2: design strategy (techniques by complexity + target-model adaptation)";
    "Phase 2: design strategy (techniques by complexity + target-model adaptation)" -> "Target model?";
    "Target model?" -> "Claude 5 default -> claude-5-guide";
    "Target model?" -> "Claude 4 -> claude-4-guide";
    "Target model?" -> "GPT-6 / GPT-5.x (OpenAI) -> openai-gpt-guide";
    "Target model?" -> "GLM 4.7 -> glm-47-guide";
    "Target model?" -> "Gemini 3 -> gemini-3-guide (Deep Research -> gemini-3-deep-research-guide)";
    "Claude 5 default -> claude-5-guide" -> "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?";
    "Claude 4 -> claude-4-guide" -> "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?";
    "GPT-6 / GPT-5.x (OpenAI) -> openai-gpt-guide" -> "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?";
    "GLM 4.7 -> glm-47-guide" -> "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?";
    "Gemini 3 -> gemini-3-guide (Deep Research -> gemini-3-deep-research-guide)" -> "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?";
    "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?" -> "Phase 3: deliver the artifact (format by type, recipient, and platform)" [label="person: full wrapper, usage + testing"];
    "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?" -> "Phase 3: deliver the artifact (format by type, recipient, and platform)" [label="process: body only, no wrapper"];
    "Recipient: a person, a process a workflow dispatches it to, or the invoking session itself?" -> "Phase 3: deliver the artifact (format by type, recipient, and platform)" [label="invoking session: ruleset only"];
}
```

### Phase 1: Prompt Scoping (Traditional Prompts Only)

Before generating the prompt, understand its intended purpose. Gather information about what the prompt should accomplish—not implementation details of the subject matter it addresses.

**Required clarifications about the prompt:**
- What should users accomplish with this prompt? (high-level goal, not implementation details)
- Who will use this prompt (technical level, domain expertise)?
- What makes the prompt successful (output quality, format, completeness)?
- Target platform: Claude Web, Claude Desktop, or API?
- Recipient: a person who copies the artifact, or a process an invoking workflow dispatches it to? Take it from context — a request naming a worker, subagent, CLI, or dispatch path is a process; ask only when neither reading fits.
- Target model: Claude 5 (default; Opus 5.5 recommended), Claude 4, GPT-6 or GPT-5.6 (OpenAI), GLM 4.7, or Gemini 3? (only ask if user mentions Claude 4, GPT, OpenAI, GLM, Z.ai, Gemini, or model adaptation)

**Clarify prompt ambiguities (stay at the prompt level, don't dive into subject matter):**
- If variations might be beneficial, ask if user wants alternative prompt approaches
- If the prompt's scope is unclear, confirm what it should and shouldn't address
- If success criteria are vague, propose concrete metrics for the prompt's output

### Refinement Mode

When the user is modifying a prompt that was previously generated in this conversation:

**Detection signals:**
- References to "the prompt", "that prompt", "the previous prompt"
- Modification requests: "make it more...", "adjust...", "change the...", "add..."
- Feedback on results: "it didn't work because...", "the output was too..."

**Streamlined workflow:**
- Skip full requirements gathering - the context is already established
- Ask a targeted question: "What specifically should change?"
- Focus on the delta, not full re-specification
- Preserve unchanged aspects of the original prompt
- Deliver the complete refined prompt (not just the changes)

### Ruleset Mode

When an invoking session asks for the distilled prompt-authoring rules it will write to a session file and re-read at each later authoring step:

- Resolve the target model, the dispatch path, and the artifact types the session will author; ask when any of the three is unstated.
- Derive the rules from that target model's guide, and state each as a directive the session applies without reloading the guide.
- Deliver the ruleset per `references/output-formats.md#14-session-scoped-rulesets`.

### Phase 2: Design Strategy

Select appropriate techniques based on task complexity:

**For simple tasks:**
- Clear, direct instructions
- Explicit output format specification
- Relevant examples if format is critical

**For complex tasks:**
- Chain of thought prompting with XML structure
  → See `references/techniques/chain-of-thought.md`
- Multi-shot examples for consistency
  → See `references/techniques/multishot-prompting.md`
- Prompt chaining for multi-step workflows
  → Technique: `references/techniques/prompt-chaining.md`; ready template: `examples/prompt-chain-template.md`

**For optimization:**
- Analyze current prompt structure and gaps
- Identify specific failure modes
- Remove personas and "You are …" identity openers: state the task directly and convert whatever they implied (tone, depth, audience) into explicit requirements — keep a persona only where it is the requested output (character roleplay)
- Apply targeted improvements with documented rationale; document the before/after with `examples/optimization-report.md`

### Phase 3: Delivery

Deliver the artifact as a markdown block optimized for its recipient and target platform. Choose the structure from the [Output Format](#output-format) table below (details in `references/output-formats.md`); for a system prompt, adapt `examples/system-prompt-template.md`.

**Person as recipient (Claude Web / Claude Desktop) — default:**
- Complete prompt in a single code block
- No API parameters unless requested
- Include usage instructions and testing suggestions

**Process as recipient (an invoking workflow dispatches the artifact):**
- Artifact body only — no wrapper, no usage instructions, no testing suggestions, and no API parameters the dispatch path cannot take
- Every fact the recipient needs stated inline, and the caller's block names, order, and output contract kept verbatim
- Details: `references/output-formats.md#13-dispatched-prompts`

**API format (only when explicitly requested):**
- Include max_tokens and effort recommendations; temperature only for Claude 4 / earlier (Claude 5 rejects non-default temperature)
- Separate system and user message components
- Provide JSON structure if needed

## Essential Techniques Reference

### Be Clear and Direct
- Provide contextual information (purpose, audience, workflow)
- Use numbered steps for sequential instructions
- Specify exact output format requirements
- Tell Claude what TO do, not what NOT to do
→ Deep dive: `references/techniques/be-clear-and-direct.md`

### Use XML Tags
- Separate prompt components: `<instructions>`, `<context>`, `<examples>`
- Structure outputs: `<thinking>`, `<answer>`, `<analysis>`
- Nest tags for hierarchical content
- Be consistent with tag naming
→ Deep dive: `references/techniques/xml-tags.md`

### Chain of Thought
- Basic: "Think step-by-step"
- Guided: Outline specific thinking steps
- Structured: Use `<thinking>` and `<answer>` tags
- Use for complex reasoning, analysis, or multi-step tasks
- Claude 5: thinking is adaptive and on by default; the `<thinking>`/`<answer>` scaffolding and manual `budget_tokens` are Claude 4 / earlier patterns (see `references/claude-5-guide.md#adaptive-thinking`)
→ Deep dive: `references/techniques/chain-of-thought.md`

### Multishot Prompting
- Include 3-5 diverse, relevant examples
- Wrap in `<examples>` tags with nested `<example>` tags
- Cover edge cases and variations
- Ensure examples match desired output format exactly
→ Deep dive: `references/techniques/multishot-prompting.md`

### Roles (Do Not Use)
- Do not write roles or personas into prompts
- "You are …" identity openers add nothing — state the task and constraints directly
- State tone, format, length, and audience as explicit output requirements instead
- For domain accuracy, provide domain context and reference material
- For a process recipient, state its authority and boundaries — act or report, what it may write, which tools it may use; that is capability context, not a persona
- Sole exception: character roleplay, where the persona is the requested output
→ Deep dive: `references/techniques/system-prompts-and-roles.md`

### Prefilling (API only — Claude 4 and earlier)
- Start assistant response to enforce format
- Skip preambles by prefilling `{` for JSON
- Maintain character in roleplay scenarios
- Cannot end with trailing whitespace
- Claude 5: prefill returns a 400 error — use Structured Outputs or a direct "respond without preamble" instruction instead
→ Deep dive: `references/techniques/prefilling.md`

## Model-Generation Optimizations

Claude 5 (Opus 5.5, Opus 5, Sonnet 5, Fable 5.1) is the default target; Opus 5.5 is the recommended default model. Claude 4 (Opus 4.x, Sonnet 4.x) and Haiku 4.5 prompts often over-steer Claude 5 — re-tune when migrating.

### Claude 5 (default)

Claude 5 exercises more judgment and needs less scaffolding. Steer with the `effort` parameter and targeted, positive instructions.

**Control length on Opus 5** — Opus 5 runs longer by default and effort doesn't reliably change visible length, so prompt for concision (Sonnet 5 is already concise; Fable 5 can over-elaborate at high effort):
```
Provide concise, focused responses. Skip non-essential context, and keep examples minimal.
```

**Remove carried-over verification on Opus 5** — Opus 5 self-verifies and self-corrects, so "double-check your answer", "add a verification step", and "use a subagent to verify" cause over-verification. Delete them. Do not generalize this: Fable 5 long-horizon runs keep verification explicit with fresh-context verifier subagents, and Sonnet 5 has no finding either way — leave its verification instructions alone (see the guide).

**Guide subagent delegation** — Claude 5 spawns subagents without being told to; give explicit criteria. Opus 5 and Fable 5 delegate more readily than prior models (Sonnet 5 has no such finding). Cap spawn counts for cost-sensitive Opus 5 work; Fable 5 is built to delegate freely (see the guide):
```
Delegate to a subagent only for large, genuinely independent, parallelizable tasks.
Keep spawn counts low.
```

**Soften aggressive language** — `CRITICAL: You MUST...` over-triggers; use "Use ... when ...".

**Breaking changes** — assistant prefill, `budget_tokens`, and non-default `temperature`/`top_p`/`top_k` return a 400 error; on Opus 5.5 and Fable 5.1 so does forced `tool_choice` (`any` / `tool`). Use adaptive thinking with `effort`, Structured Outputs instead of prefill, and `tool_choice: auto` plus strict tool use instead of forced tool choice.

→ Full guide, per-model specifics (Opus 5.5 / Opus 5 / Sonnet 5 / Fable 5), the Opus 5 → Opus 5.5 checklist, and Claude 4 → 5 migration: `references/claude-5-guide.md`

### Claude 4 and earlier

Claude 4 needs "above and beyond" behavior requested explicitly:
- Request thoroughness ("Include as many relevant features and interactions as possible. Go beyond the basics.").
- Explain *why* an instruction matters — Claude 4 generalizes from the reason.
- Add anti-reward-hacking wording for coding ("high-quality, general-purpose solution; don't hard-code test cases; say so if the task is unreasonable").
- Prompt reflection on tool results before the next step.

→ Full guide: `references/claude-4-guide.md`

## Claude 4 → Claude 5 Migration (When Requested)

When the user asks to migrate, re-tune, or update existing Claude 4 content — a prompt, skill, agent, or rules file — for Claude 5: diagnose and remove the carried-over Claude 4 scaffolding that now over-steers, rather than rewriting from scratch. Apply the re-tuning moves from the Claude 5 sections above — remove verification scaffolding when the target is Opus 5 (Fable 5 long-horizon runs keep it; Sonnet 5 keeps what it has), soften aggressive triggers, give and cap subagent-delegation criteria, replace the 400-error API parameters (assistant prefill, `budget_tokens`, non-default `temperature`/`top_p`/`top_k`) — plus, for LLM-targeted content, the "less is more" optimization principle.

Deliver a before/after so each change and its rationale stay visible: `references/output-formats.md#10-claude-4-to-claude-5-migrations`. Worked transformations: `examples/claude-4-to-5-migration.md`. Full per-model detail and the Claude 4 → 5 migration reference: `references/claude-5-guide.md`.

## GPT-6 / GPT-5.6 (OpenAI) Adaptation (When Requested)

When the user explicitly targets GPT-6 (Astra, Sol, Luna), GPT-5.6 (Sol, Terra, Luna), or another OpenAI GPT-5-family model: define the outcome, constraints, evidence, and completion bar, then leave the path to the model.

- Write outcome-first prompts with explicit success criteria and stop rules; remove step scripts.
- State each instruction once and remove contradictions — conflicting rules destabilize GPT-5.6 more than missing detail.
- Reserve ALWAYS/NEVER/MUST for true invariants; give decision rules for judgment calls.
- Re-check blanket brevity instructions (GPT-5.6 is concise by default); set `text.verbosity` for the default level of detail.
- Define autonomy and approval boundaries (report vs. act vs. confirm) instead of repeating "ask first".
- API: Responses API, `reasoning.effort` (`none`–`max`, default `medium`; GPT-6 Astra has no `none`), `reasoning.mode: "pro"` for quality-first single answers.
- GPT-6: choose Astra, Sol, or Luna by workload (no Terra; Astra is the top tier); state that action requests authorize the work and that the user's instructions take precedence over skill guidelines; make instruction-file guidance conditional rather than blanket; bound re-verification; keep the confirmation tier for external writes and destructive actions. Evaluate each pattern per model — the guidance was observed on Astra.

→ Full guide, API parameters, per-tier selection, GPT-6 and GPT-5.x migration, and pre-5.6 baseline: `references/openai-gpt-guide.md`; transformation examples: `examples/openai-gpt-adaptation.md`

## GLM 4.7 Adaptation (When Requested)

When the user explicitly targets GLM 4.7: it treats polite, buried instructions as optional, so make directives firm and front-loaded.

- Front-load all mandatory rules in the first 200 words.
- Convert soft language to hard directives (MUST / ALWAYS / NEVER).
- Add explicit output templates and a FORBIDDEN-patterns block to block generic responses.
- Add a self-verification block; force language with `ALWAYS respond in English`.
- API: enable thinking, temperature 0.6–0.7, and GLM stop tokens.

→ Full adaptations, patterns, and API config: `references/glm-47-guide.md`; transformation examples: `examples/glm-47-adaptation.md`

## Gemini 3 Adaptation (When Requested)

When the user explicitly targets Gemini 3: it defaults to minimal output and handles long context differently than Claude.

**Deep Research mode** — if the user mentions "deep research", citation-backed literature reviews, or extended autonomous web research, use `references/gemini-3-deep-research-guide.md` and `examples/gemini-3-deep-research.md` (autonomous 5–60 min runs, plan review, ~10% citation-error, explicit scope/temporal/source constraints). Otherwise apply standard adaptations:

- Keep temperature at 1.0 — changing it causes looping or degraded output.
- Place instructions after long context; put constraints at the prompt end.
- Request verbosity explicitly and always include 2–3 few-shot examples.
- Control format with response-prefix strings; add a self-verification block.
- Flash (routine, high-volume) vs Pro (complex) — identical prompting.

→ Full adaptations, model selection, and API config: `references/gemini-3-guide.md`; transformation examples: `examples/gemini-3-adaptation.md`

## Output Format

Select the appropriate format based on artifact type:

| Type | Format | Reference |
|------|--------|-----------|
| Traditional prompts | Full wrapper (Purpose, Best Used For, etc.) | `references/output-formats.md#1-traditional-prompts` |
| LLM-targeted content | Optimized content only, no wrapper | `references/output-formats.md#2-llm-targeted-content` |
| Refinements | Add "Changes Made" section | `references/output-formats.md#3-refinements` |
| Prompt chains | Chain Overview with steps | `references/output-formats.md#4-prompt-chains` |
| GLM 4.7 prompts | Add API Configuration | `references/output-formats.md#5-glm-47-prompts` |
| Claude-to-GLM adaptations | Before/After comparison | `references/output-formats.md#6-claude-to-glm-adaptations` |
| Gemini 3 prompts | Add API Configuration | `references/output-formats.md#7-gemini-3-prompts` |
| Claude-to-Gemini adaptations | Before/After comparison | `references/output-formats.md#8-claude-to-gemini-adaptations` |
| Gemini Deep Research prompts | Add Deep Research Notes + Iteration Suggestions | `references/output-formats.md#9-gemini-deep-research-prompts` |
| Claude 4 → Claude 5 migrations | Before/After comparison | `references/output-formats.md#10-claude-4-to-claude-5-migrations` |
| GPT-6 and GPT-5.6 prompts | Add API Configuration | `references/output-formats.md#11-gpt-6-and-gpt-56-prompts` |
| Claude-to-GPT adaptations (GPT-6, GPT-5.6) | Before/After comparison | `references/output-formats.md#12-claude-to-gpt-adaptations-gpt-6-gpt-56` |
| Dispatched prompts (process recipient) | Prompt body only, no wrapper or usage sections | `references/output-formats.md#13-dispatched-prompts` |
| Session-scoped rulesets | Rules only, no wrapper and no prompt | `references/output-formats.md#14-session-scoped-rulesets` |

## Quality Checklist

Before delivering any artifact, verify:
- [ ] Instructions are unambiguous and complete
- [ ] Proper use of XML tags and formatting
- [ ] Relevant examples included where helpful
- [ ] Clear success criteria provided
- [ ] Handles edge cases appropriately
- [ ] No personas and no "You are …" identity openers — task, tone, depth, and audience stated as explicit requirements (exception: character roleplay)

## When to Ask for Clarification

For **traditional prompts**, ask the user for clarification when:
- Use case is ambiguous or too broad
- Multiple valid approaches exist
- Output format preferences are unclear
- User might benefit from prompt variations
- Success criteria need definition
- Target platform is not specified

Example clarification:
```
Before I create this prompt, I have a few questions:
- Should this prompt handle [specific edge case]?
- Do you want the output in [format A] or [format B]?
- Would you like me to provide alternative variations?
```

## Specialized Domains

### Software Development
- Code generation: completeness, error handling, best practices
  → See `references/prompt-patterns/code-generation.md`
- Code review: structured criteria, actionable feedback
  → See `references/prompt-patterns/code-review.md`
- Architecture: systematic exploration, trade-off analysis
  → See `references/prompt-patterns/comparative-analysis.md`
- Debugging: methodical problem-solving, hypothesis testing
  → See `references/prompt-patterns/root-cause-analysis.md`

### Business & Analysis
- Data analysis: clear insights, visualization recommendations
  → See `references/prompt-patterns/structured-analysis.md`
- Report generation: professional formatting, executive summaries
  → See `references/prompt-patterns/chain-patterns.md`
- Decision support: structured options, risk assessment
  → See `references/prompt-patterns/comparative-analysis.md`

### Creative & Content
- Writing: tone consistency, audience adaptation
  → See `references/prompt-patterns/structured-content-generation.md`
- Documentation: technical accuracy, user-friendliness
  → See `references/prompt-patterns/documentation-generation.md`
- Marketing: brand voice, conversion optimization
  → See `references/prompt-patterns/text-rewriting.md`
Pliki w paczce (43)
  • .claude-web 1 KB
  • SKILL.md 26 KB
  • examples/claude-4-to-5-migration.md 5 KB
  • examples/gemini-3-adaptation.md 10 KB
  • examples/gemini-3-deep-research.md 8 KB
  • examples/glm-47-adaptation.md 13 KB
  • examples/openai-gpt-adaptation.md 8 KB
  • examples/optimization-report.md 7 KB
  • examples/prompt-chain-template.md 5 KB
  • examples/system-prompt-template.md 4 KB
  • references/claude-4-guide.md 10 KB
  • references/claude-5-guide.md 17 KB
  • references/gemini-3-deep-research-guide.md 8 KB
  • references/gemini-3-guide.md 12 KB
  • references/glm-47-guide.md 19 KB
  • references/openai-gpt-guide.md 21 KB
  • references/output-formats.md 8 KB
  • references/prompt-patterns/README.md 1 KB
  • references/prompt-patterns/api-design.md 1 KB
  • references/prompt-patterns/chain-patterns.md 2 KB
  • references/prompt-patterns/code-generation.md 1 KB
  • references/prompt-patterns/code-review.md 1 KB
  • references/prompt-patterns/comparative-analysis.md 1 KB
  • references/prompt-patterns/customer-feedback-analysis.md 1 KB
  • references/prompt-patterns/data-transformation.md 1 KB
  • references/prompt-patterns/document-review.md 1 KB
  • references/prompt-patterns/documentation-generation.md 1 KB
  • references/prompt-patterns/pattern-customization.md 1 KB
  • references/prompt-patterns/root-cause-analysis.md 1 KB
  • references/prompt-patterns/structured-analysis.md 1 KB
  • references/prompt-patterns/structured-content-generation.md 1 KB
  • references/prompt-patterns/test-case-generation.md 1 KB
  • references/prompt-patterns/text-rewriting.md 1 KB
  • references/techniques/README.md 1 KB
  • references/techniques/be-clear-and-direct.md 2 KB
  • references/techniques/chain-of-thought.md 2 KB
  • references/techniques/long-context.md 1 KB
  • references/techniques/multishot-prompting.md 1 KB
  • references/techniques/prefilling.md 1 KB
  • references/techniques/prompt-chaining.md 1 KB
  • references/techniques/reduce-hallucinations.md 1 KB
  • references/techniques/system-prompts-and-roles.md 2 KB
  • references/techniques/xml-tags.md 1 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · it-bens/ai-tools100MIT
Muvon/octomind-tap100Apache-2.0
TerminalSkills/skills100Apache-2.0
ericrisco/rsc-harness100MIT
wrongstack/wrongstack100MIT
rootcastleco/rei-skills100MIT
FrancoStino/opencode-skills-collection100MIT
davila7/claude-code-templates100MIT
CodeAlive-AI/ai-driven-development100MIT

Migawka bazy z 13.11.2026 · wersja paczki 78479628f620 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 100prompt-engineer
    Pomaga pisać i recenzować prompty oraz instrukcje dla systemów AI, w tym system prompty, i poprawiać istniejące instrukcje dla Claude, GPT i innych LLM.
  • 100system-prompt-creator
    Tworzy wysokiej jakości prompty systemowe dopasowane do modelu, dla Claude, GPT, Gemini i modeli otwartych, oraz poprawia istniejące instrukcje, definiując zachowanie asystenta, chatbota lub agenta.
  • 86prompt-optimizer
    Poprawia i optymalizuje prompty: przepisuje, buduje szablony i prompty systemowe, chain of thought i few-shot, dla Claude i GPT, z biblioteką promptów, w 20 funkcjach.
  • 100prompt-optimizer
    Optymalizuje prompty i prompty systemowe dla najnowszych modeli Claude (rodzina Claude 5 oraz Opus 4.8, Sonnet 4.6, Haiku 4.5) według oficjalnych wskazówek Anthropic dotyczących inżynierii promptów.
  • 100prompt-engineering-patterns
    Pomaga optymalizować prompty, poprawiać ich wydajność, projektować szablony, debugować problemy oraz stosować chain-of-thought, ustrukturyzowane i few-shot promptowanie w produkcyjnych aplikacjach LLM.
  • 100prompt-engineering-patterns
    Pomaga optymalizować prompty, poprawiać ich wydajność, projektować szablony, debugować problemy oraz stosować chain-of-thought, ustrukturyzowane i few-shot promptowanie w produkcyjnych aplikacjach LLM.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.