Ask for tables and JSON without breaking your workflow

Goal: Ask for tables or JSON with a frozen schema; unknown stays null—no pretty fake fills.

Best for: Briefs, matrices, checklists, light automation handoffs.

Structured output fails when you underspecify keys or allow prose wrappers. Name columns/keys. Ban extra commentary. Validate JSON before you pipe it anywhere.

Ask for structure when you will paste into a sheet or script. Give a schema, one tiny example, and what to do if a field is unknown (null/TBD—not invented).

Why beginners skip this

Asking only “make a table” with no schema lets the model invent column names—downstream parse will break. Write the fields first.

Step 1 — Table ask

Practice “table only”: use a real list you have; do not invent rows.

Return ONLY a markdown table with columns: [A | B | C].
No intro paragraph. Unknown allowed. Do not invent values.

Step 2 — JSON ask

Then JSON: freeze the schema first, paste notes in; unknown fields stay null.

Return ONLY valid JSON with keys: title (string), bullets (array of 3 strings), risks (array of 2 strings).
No markdown fences if the app wraps JSON for you—otherwise one fence.
If unknown, use null or "Unknown"—do not invent.

Extract meeting actions into JSON with fields: owner, task, due (YYYY-MM-DD or null), source_quote. If due unknown use null—do not invent dates. Provide one example object then parse this note: [paste].

Step 3 — Validate

Paste JSON into a validator or JSON.parse in a scratchpad. Tables: paste into Sheets and check column counts.

Scene: Sheets without tears

The model returns ONLY the table. You paste into Sheets; columns line up. No essay wrapper to delete. Unknown cells stay empty on purpose.

Refuse pretty invalid JSON

If the model wraps JSON in commentary, send it back. Pretty-but-invalid breaks pipelines. Prefer boring valid output.

Your last reply was not valid JSON. Return ONLY the JSON object with keys [list]. No commentary.

From messy meeting notes, output JSON array of decisions: {decision, owner|null, due|null, quote}.
If missing → null, never invent. Return JSON only + one Unknown note.

Save this card

Card: Want structure—give a schema; unknown stays unknown—no pretty fake values. Keep this beside the tool; read before you start.

Before you start

Keep the last failing output visible. You will repair the prompt for “Ask for tables and JSON without” against that failure—not a fresh toy ask. Guide: prompt-structured-output-json-tables. Keep the real job visible the whole time; a blank box means stop and fetch the artifact first.

How to use this week

This week for a real task write a JSON/table schema: field names + types + required; no run without the schema.

Keep it real

Run this guide’s Practice now on work you will actually ship this week—not a fictional scenario. Dated notes beat perfect theory.

Ship the practice

Close the tab only after Practice now is done on a real artifact.

Do not bookmark-only

Finish Practice now, then tick this guide on your path. Titles unread are not skills.

Close the loop

Write one line: next time I run with no schema, I refuse.

Check the unknown path

What should the model do when a field is missing—null, TBD, or ask you? Write it once in Boundaries.

From notes, JSON array: {decision, owner|null, due|null, quote}.
Missing → null. JSON only + Unknown note. No invented owners.
Valid JSON only; unknown → null; no trailing commentary.

Red flag

Pretty fake values for unknowns; no schema—require unknown to stay unknown.

Pitfalls

Temptation: skip the habit for prompt-structured-output-json-tables to save five minutes.
Job still: Get markdown tables or simple JSON you can
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.

Practice now

Null policy: define how nulls appear—never silent omit (Asia/Shanghai).

Invalid reject: feed one invalid row; model must refuse (Asia/Shanghai).

Schema freeze: freeze schema before generating rows (Asia/Shanghai).

Structured drill: one JSON/table schema; reject free prose that breaks it (Asia/Shanghai).

Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.

Tonight write a three-line schema (name+type+required) for real material, then generate JSON/table.

Habit stamp: after one live pass, write the refuse line you actually used on the Save card and date it (Asia/Shanghai).

Today card (copy) — prompt-structured-output-json-tables: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

Extra pass: do the card on a real file today—reading alone does not count.

One sentence takeaway

Field names + types + required; no run without the schema.

Next

Claim–evidence matrix · GFB · Rubrics

要表格与 JSON，且不弄乱工作流

目标： 拿到能直接贴进文档或表格的 markdown 表，或简单 JSON。不编造字段，也不产出坏语法。

适合： 简报、矩阵、清单，以及要交给轻度自动化的交接。

结构化输出常在两处翻车：键没写清，或允许散文把 JSON 包起来。请写清列名和键名，并禁止额外评论；接到任何流水线前，先校验。

真要贴进表格或脚本时，就明说要结构：给 schema，附一个小例子，并写清字段未知时怎么办（null 或 TBD，不要编造）。

为什么新手跳过

只要「整理成表」不写 schema，模型会自由列名；下游解析会炸。先写字段。

步骤 1——要表

先练「我只要一张表」：材料用你手头真实清单，别编行。

只返回 markdown 表，列名为：[A | B | C]。
不要写开场段落。允许填 Unknown。不要编造值。

步骤 2——要 JSON

再练 JSON：先把 schema 写死，再往里贴笔记；未知用 null。

只返回合法 JSON，键：title（字符串）、bullets（3 个字符串的数组）、risks（2 个字符串的数组）。
如果下游应用自己会包 JSON，就不要再套 markdown 围栏；否则只用一个围栏。
未知就用 null 或 "Unknown"：不要编。

把会议行动项抽成 JSON，字段：owner、task、due（YYYY-MM-DD 或 null）、source_quote。到期日未知就用 null：不要编日期。先给一个示例对象，再解析下面这条笔记：【粘贴】。

步骤 3——校验

把 JSON 丢进校验器，或在草稿里跑一次 JSON.parse。如果是表格，贴进 Sheets，看列数齐不齐。

场景：不掉眼泪的 Sheets

理想情况：模型只返回表格；你贴进 Sheets，列正好对齐；没有要删的散文外壳；Unknown 格故意留空。

拒绝好看但非法的 JSON

如果模型用评论把 JSON 包起来，打回去。好看但非法会弄坏流水线。宁可要无聊的合法输出。

你上一条不是合法 JSON。只返回键为 [列表] 的 JSON 对象。不要评论。

模型返回几乎是 JSON，但尾部拖着评论。请写一条修复提示词：只返回符合 schema 的合法 JSON；未知用 null；禁止解释性尾巴。
边界：不编造；未知标 [?]；发送前停下。
完成检查：一分钟内能判过/不过。

保存这张卡

记下：今天改管用的那条提示词（含边界）+ 版本或日期。

开始前

调试时把上一次失败输出留在眼前。

本周怎么用

本周对一个真任务写 JSON/表头 schema：字段名 + 类型 + 必填；缺 schema 不跑。

保持真实

用本周真要交付的工作做「现在就练一次」，别编假场景。带日期的实操笔记，比完美理论有用。

把练习做完

用真实材料做完「现在就练一次」，再关标签页。

不要只收藏

先做完「现在就练一次」，再在学习路径上打勾。只收藏不练，标题变不成技能。

闭环

写下一行：下次缺 schema 就开跑，我拒绝。

检查未知路径

字段缺失时模型该怎么做：填 null、写 TBD，还是问你？在边界里写清楚一次。

红旗

只要「整理成表」不写 schema，模型会自由列名，下游解析会炸；先写字段。

常见坑

现在就练一次

今晚就练：对一个真任务写 JSON/表头 schema：字段名 + 类型 + 必填；缺 schema 不跑。

一句话带走

字段名 + 类型 + 必填；缺 schema 不跑。

下一步

主张–证据矩阵 · 目标·格式·边界 · 评分表