Quality Bar Contract

贴在任何请求之前的简短「质量契约」：先指名世界级参照物、写出评分表、禁止含糊取值，再自评打分直到各项不低于 9 分。

本库中最短的提示词。它把「做好一点」换成可校验的目标：为该任务指名 2-3 个真实的世界级参照、写出 5-7 个维度的评分表、凡该给具体数值处一律禁止写「合适的值」，然后以敌意专家身份自评并持续修订，直到每个维度都不低于 9 分。文中还给出微调语句，以及不值得付出这份开销的场景。

# Quality Bar

**Use when:** any task where "it works" is not good enough.
**Works with:** Claude Code, Codex, Cursor, ChatGPT, Gemini, any chat model.
**Pairs with:** [loop_engineering.md](loop_engineering.md), [self_critique.md](self_critique.md)

This is the shortest thing in the library that changes output the most. Paste it above your real request.

---

## The prompt

```text
QUALITY CONTRACT

Your first working answer is a draft. It is never the deliverable.

The target is not "correct" and not "good enough". The target is:
a specialist who does only this for a living would look at it and find
nothing they would change.

Before you produce anything:

1. NAME THE BAR
   Name 2 or 3 real products, libraries, papers, or codebases that are
   world class at this exact task. For each, say in one line what
   specifically makes it world class. You are now competing with those.

2. WRITE THE RUBRIC
   List 5 to 7 dimensions this work will be judged on. For each dimension
   write one sentence describing what a 10 out of 10 looks like here.
   Dimensions must be specific to this task, not generic.

3. STATE THE HARD PARTS
   List the 3 things most likely to be done badly here. These are where
   you spend your effort. Everything else is table stakes.

Then produce the work.

After you produce it:

4. SCORE IT
   Score every rubric dimension 1 to 10. Every score needs one line of
   evidence. Evidence means a specific part of your output, not a feeling.

5. GATE
   If any dimension is below 9, you are not done. Fix the lowest one and
   rescore. Repeat until every dimension is 9 or above, or until you can
   explain exactly why 9 is impossible here.

FORBIDDEN IN THE FINAL OUTPUT
- "TODO", "placeholder", "example only", "in a real app you would"
- Simplified versions offered instead of the real thing
- Any claim about behaviour you did not verify
- Any value left vague where a real value belongs
- Asking me a question you could answer yourself by looking

DETAIL FLOOR
Where a real value belongs, give the real value. Exact colour, exact
spacing, exact duration, exact type name, exact file path, exact command.
"Some padding" and "an appropriate value" are failures.

EFFORT
This is not a task to finish quickly. Use your full reasoning budget.
Thinking longer is cheaper than me finding the flaw later.

REPORTING
Do the deep work, then report briefly. I want the short answer on top and
the reasoning available underneath, not a wall of process.
```

---

## Short version

When you only have one line to spare:

```text
Your first answer is a draft. Name what world class looks like for this,
write a rubric, then produce work that scores 9+ on every dimension. No
placeholders, no vague values, no unverified claims. Report briefly.
```

---

## Why it works

Language models optimise for a plausible answer, not a best answer. The default behaviour is to stop as soon as the response would satisfy an average reader. Each clause here removes one way of stopping early.

- **Naming the bar** replaces a vague target with a concrete one. "Good button" is unmeasurable. "As good as the primary button in Linear" is measurable.
- **Writing the rubric before the work** stops the model from grading itself on whatever it happened to do well.
- **Naming the hard parts** moves effort to where it matters instead of spreading it evenly.
- **The 9 gate** turns self review from a formality into a loop with an exit condition.
- **The forbidden list** blocks the specific escape hatches models reach for under pressure.
- **The detail floor** is the single highest value line for design and UI work.

---

## Tuning

- Raise the gate to 9.5 for work that ships to users. Drop it to 8 for throwaway scripts.
- Replace step 1 with your own named references when you have them. "Match the API design of the Stripe Node SDK" beats anything the model picks on its own.
- If the model starts padding the output to look thorough, add: `Length is not quality. Cut anything that does not carry information.`
- If the model spends the whole response on the rubric and little on the work, add: `Keep steps 1 to 3 under 150 words total.`

---

## Do not use when

- The task is genuinely trivial and you want a fast answer.
- You are exploring and want cheap throwaway options, not a finished artefact.

Overhead is real. This contract costs tokens and time. Spend it where the output matters.