tflow-prompt
Use when the user shares a raw, vague, or first-draft prompt and wants it clarified, tightened, strengthened, or made more reliable — or asks to improve, rewrite, or optimize a prompt, write a system prompt, or apply prompt-engineering techniques like few-shot examples, chain-of-thought, role framing, or XML structuring
- License 许可
- MIT
- 所属仓库
- terra0rg/tflow
- 最近更新
- 2026-09-09
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- tflow-prompt
- license
- MIT
- description
- Use when the user shares a raw, vague, or first-draft prompt and wants it clarified, tightened, strengthened, or made more reliable — or asks to improve, rewrite, or optimize a prompt, write a system prompt, or apply prompt-engineering techniques like few-shot examples, chain-of-thought, role framing, or XML structuring
- compatibility
- Portable Agent Skill source for Claude Code, OpenAI Codex, and runtimes that support SKILL.md; requires no scripts, tools, or network access.
02 资源文件 3
- 参考文档evals/evals.json2.4 KB
- 参考文档references/examples.md2.7 KB
- 脚本tests/run-tests.sh3.8 KB
03 SKILL.md 正文
tflow Prompt
Use this skill to turn a user's base prompt into a stronger one. The default is a single pass: read the prompt, apply the techniques that actually help it, and hand back the rewritten prompt plus a short change log. Always rewrite and explain — never return a silently transformed prompt the user cannot audit.
Enhance to earn it, not to decorate. Add a technique only when it makes the prompt measurably clearer for its goal; if the original is already tight, say so and change little. Match effort to the prompt's complexity — a one-liner usually needs a clarity pass, not all seven techniques.
Prompt-engineering specifics drift between model releases: token limits, provider flags, and model names change. Treat the technique principles here as durable, but tell the user to verify any concrete model name, token limit, or API field against the provider's current documentation rather than trusting a fixed value.
Locating the prompt
Take the prompt to enhance from the user's latest message or from earlier in the conversation. If no candidate prompt is present, ask the user to paste the one they want strengthened before going further.
Process
- Infer the goal. What outcome does the user want, for what audience, in what shape? If the prompt is too thin to infer a goal, ask up to three targeted questions (typically goal, audience, result shape), then proceed.
- Run the completeness check. A well-formed prompt covers up to four components: an instruction (the task), context (background and motivation), input data (the material to act on), and an output indicator (the shape of the result). Note only the ones that are missing and matter for this task — not every prompt needs all four.
- Apply the techniques in order, each only where it helps. The order in Techniques is deliberate: clarity first, structure and examples next, role and reasoning last.
- Scale to complexity. Reserve examples, structure, role framing, and step-by-step reasoning for prompts whose difficulty earns them. Name anything you remove as noise, not just what you add.
- Return the two-part result below.
Techniques, in order
- Clarity & directness — state the task and its constraints explicitly; replace vague language with concrete instructions.
- Context & motivation — explain why, so the model generalizes beyond the literal words.
- Examples — show 2–5 diverse examples for format-sensitive or nuanced tasks; demonstrate the desired output rather than only describing it.
- Structure — separate instruction, context, and input into labeled sections (headers or tags) so they are not confused for one another.
- Role — give the model a role or expertise level when tone or domain rigor matters.
- Reasoning — invite step-by-step thinking for analytical or multi-factor tasks, before the final answer.
- Decomposition — split complex, multi-stage work into sequential prompts where each output feeds the next.
Worked before/after rewrites are in examples.
Result format
Return exactly two parts, in this order:
- Enhanced prompt — the rewritten prompt in a fenced code block, ready to copy and use verbatim.
- Change log — a short bullet list, each line tagged by the technique that
motivated it and a one-line reason, ending with a completeness note: one
line naming which of the four components (instruction, context, input,
output indicator) are now present, and any the user must still supply — for
example real input data or domain facts. Use these tags:
clarity— sharpened or disambiguated the instruction.context— added background or motivation.example— added or restructured examples.structure— introduced or tidied sectioning.role— set or adjusted a role.reasoning— invited step-by-step thinking.decomposition— split into stages or sequenced prompts.cut— removed noise, redundancy, or over-specification.
Boundaries
- This skill improves prompts; it does not run, benchmark, or evaluate them. Suggest the user test against their own success criteria, but keep execution out of scope.
- It assumes no specific provider or model. The techniques are general; defer vendor-specific detail to the provider's live documentation.
- It never fabricates the user's domain facts or input data. Missing material is named in the completeness note, not guessed.
References
- examples — worked before/after enhancements showing the two-part result and proportional technique use.
04 同仓库更多 Skills
tflow-skill-test
Use when an Agent Skill needs a TDD-style test plan before authoring, or a finished skill needs the three-layer test pass — structural gate, deterministic script checks, and judged eval scenarios
tflow-skill-idea
Use when a raw Agent Skill idea needs interactive shaping into a research-ready idea brief — surfacing the core purpose with Five Whys and a human-chosen direction before any research or authoring begins
tflow-skill-factory
Use when a raw Agent Skill idea must become a tested, reviewed, and documented skill through the full eight-step pipeline; for a single phase alone, use the matching phase skill directly
pass-well-formed
Use when you need a canonical example of a well-formed skill that passes all quality checks
pass-verb-substring
Use when you need information about typically uninstalled prior art
pass-trigger-condition
Use when an observable trigger condition applies to the current state