Prompt Writing Guide
A prompt that works twice is built, not phrased. Four parts, each with one job: who is answering, what is true and what is missing, what must come back, and one passage of the output at the quality you want.
Status: v1.0 | Category: Frameworks | Price: Free
Read the product guide before you write a prompt. Run the included tools from this product folder. Keep your prompts and outputs in your own workspace.
What this framework does
- Takes a job you would otherwise describe in one sentence.
- Fills four parts: role, input, task, example.
- Assembles them into one labelled prompt that a model can follow in order.
- Checks the prompt against 14 rules and exits non-zero on a break.
- Grades the prompt part by part, names the weakest part and the repair that buys the most.
- Shows the test that matters: run it again with different facts and compare the shape.
Route
- Write a new prompt → read
GUIDE.md, thenpython3 tools/rit.py --example > brief.json, edit the four values, and runpython3 tools/rit.py brief.json --out prompt.txt. - Find wording for one part →
references/part-library.md. - Fix a failure you can see in the output →
references/failure-list.md. - Improve a thin or average example →
references/example-craft.md. - Judge a prompt you already have →
python3 tools/rit_grade.py prompt.txt.
Rules that are not negotiable
- The role names a discipline and a level of experience. It does not praise.
- The input states the situation, names at least one constraint, and says what is missing. Absent material is named, not invented.
- The task names a deliverable, its sections, its length, its register, and what the output must be usable for.
- The example is a passage of the output, not a description of it. The shape, the depth and the register, and nothing else.
- A prompt that cannot be rerun with different input is a question, not a prompt.
- The checker reads the prompt. It does not promise what the answer will be.
Files
| File | Use |
|---|---|
GUIDE.md | The full method: the four parts, the reuse test, the failure list, the voice route |
references/part-library.md | Paste-ready blocks for all four parts, five variants each |
references/failure-list.md | 24 named failures with symptom, cause and repair |
references/example-craft.md | What a good example contains, and three ways a bad one poisons a prompt |
tools/rit.py | Assembles a prompt from a JSON brief and checks it against 14 rules |
tools/rit_grade.py | Grades a prompt out of 25 per part and names the weakest part |