agentsop-output-format-by-model

# Output-Format-by-Model — 让格式服务于任务，而不是反过来 > **One-liner**: Different output formats carry different cognitive load for the model. Code-in-JSON is the canonical proof: the same model writes worse code when wrapped in a JSON tool-call than when emitted as plain text + diff. The reverse failure (asking for prose when you need a typed object) is just as common. Pick format per (task × consumer), not by reflex. --- ## 1. 何时激活 (When to activate) Activate this skill **before** committing to an output schema in any of these situations: | Trigger | Signal | |---|---| | Designing a coder-agent | "should the model return a `apply_patch` tool call or plain-text diff?" | | Adding a tool to an existing agent | "tool input has a `code` / `query` / `sql` / `regex` field — should I nest it in JSON or leave it as a string?" | | Building extraction / classification | "should I use `dspy.Predict` typed fields, Pydantic + `response_format=json_schema`, or just markdown?" | | Wiring an evaluator | "the metric needs a number — but the model also has to *reason* to produce it" | | Migrating a working prompt to "structured outputs" | someone said "let's make it safer with JSON schema" | | Tool-call harness adds latency / errors | repeated `json.JSONDecodeError`, escaping bugs, truncated outputs | **Anti-triggers** (skip this skill): - The format is fixed by an external API (e.g. you must return OpenAI function-call JSON — no choice). - One-shot exploratory prompting where no consumer parses the output yet. - The task is *itself* about JSON (e.g. "fix this malformed JSON") — see §6. --- ## 2. 核心心智模型 (Core mental model) ### 2.1 Three-layer claim ``` ┌──────────────────────────────────────────────────┐ │ FORMAT FOLLOWS FUNCTION │ │ │ │ Some formats add cognitive load to the model │ │ and measurably degrade quality on the │ │ *content* the format is supposed to wrap. │ └──────────────────────────────────────────────────┘ ▲ ▲ │ │ What's being consumed? Who consumes it? (code? prose? entities? (human reader? parser? number? action selection?) downstream LM? compiler?) │ │ └──────────┬───────────────┘ ▼ FORMAT SELECTION (text+diff | markdown | JSON | tool_use | grammar-constrained | typed field) ``` Aider's `code-in-json` benchmark is the load-bearing empirical anchor: - GPT-4 Turbo refactor benchmark with **unified diff**: **61%** pass. - Same model with **JSON-wrapped code**: **20%** pass. - That is a **3×** swing from format alone, not from model choice or prompt rewording. [aider.chat/2023/12/21/unified-diffs.html, aider.chat/2024/08/14/code-in-json.html] - The pattern holds across every major model tested in 2024 (Claude Sonnet, DeepSeek Coder, GPT-4o). Even with OpenAI strict-mode JSON validity enforced, **the code *inside* the JSON degraded** — more `SyntaxError` / `IndentationError`. Sonnet kept syntax clean but still scored lower overall. [aider.chat/2024/08/14/code-in-json.html] Academic generalization, same year: - "Let Me Speak Freely?" (arXiv 2408.02442) — **format-restricted reasoning drops 10–15%** on math and complex analysis vs free-form-then-convert. [arxiv.org/abs/2408.02442] - StructEval (arXiv 2505.20139) — even o1-mini only scores 75.58 average on structural-output generation. [arxiv.org/html/2505.20139v1] ### 2.2 The two reflexes to unlearn | Reflex | When it's wrong | |---|---| | "Structured output is always safer." | False for code, multi-step reasoning, free-form prose. Strictness ≠ quality of contents. | | "Markdown is only for humans." | False — markdown is *also* the highest-fidelity wire format for many LM-to-LM hand-offs (Aider uses it; DSPy's default chat adaptor uses field-marked markdown over JSON for many signatures). | ### 2.3 Mental model: format as a *tax* on the generation surface Every formatting requirement consumes some of the model's attention budget. The tax is: - **Low** for formats the model has seen millions of times in pretraining (markdown, unified diff, Python). - **Medium** for JSON containing simple scalar fields (entity extraction, dates, booleans). - **High** for JSON containing code, multi-line strings with escapes, or deeply nested reasoning. Each `\n` becomes `\\n`; each quote becomes `\"`; the model has to track this *while also* solving the actual problem. ⇒ Heuristic: **the more semantically dense the content, the cheaper the format must be.** --- ## 3. SOP 工作流 (Decision workflow) Three questions, in order: ### Step 1 — What's being consumed? | Content type | Default format | Why | |---|---|---| | Source code edits | **plain text + diff format** (SEARCH/REPLACE, unified diff, or `str_replace` tool with code as a *single string field*) | Aider 20%→61% on GPT-4 Turbo; same direction across all models. [aider.chat/2024/08/14/code-in-json.html] | | Source code (full file rewrite) | **plain text or markdown fenced block** | Same reason. JSON-wrap adds escaping tax. | | Structured data extraction (entities, dates, IDs, classifications) | **JSON / Pydantic / typed `OutputField`** | Schema *helps* here — fields are the task. | | Action selection (which tool to call) | **JSON tool_use** | Tool name + scalar args. The decision is structured by definition. | | Action *body* (SQL, regex, code, file contents) | **single string field inside tool_use** — do not sub-structure | Same code-in-JSON penalty applies to any code-shaped payload. | | Reasoning / intermediate steps | **markdown or Python-style scratchpad** | CoT in JSON measurably degrades; see "Let Me Speak Freely?" [arxiv.org/abs/2408.02442] | | Numeric answer with reasoning | **markdown reasoning + final answer in fenced block or final-line convention** | Don't force the reasoning into a JSON `reasoning` field — it shortens and stiffens. | | Free-form prose (summary, explanation, customer reply) | **markdown** | Native to instruction-tuned models. | | Mixed (e.g., extract entities AND rewrite the document) | **split into two calls or two passes** — see §5 Case B | One format can't serve two contents well. | ### Step 2 — Who consumes the output? | Consumer | Constraint | Implication | |---|---|---| | Human in a chat UI | Render-friendly | Markdown wins. JSON is hostile. | | `json.loads` / Pydantic parser | Must be valid | JSON with schema, or a known-safe envelope (`<result>...</result>`) with markdown body. | | Downstream LM (LM-to-LM pipeline) | Reads what was written | Markdown is *more* robust than JSON when content includes code/math; the next LM ingests it natively. | | Compiler / interpreter (PoT, code-exec sandbox) | Must be a valid program | Output **as code** in a fenced block, not as a JSON `program` field. | | Tool dispatcher (function calling) | Needs tool name + args | JSON tool_use with **scalar args**; multi-line bodies go in a single string field. | | Diff applier (Aider, git apply, str_replace) | Must apply cleanly | Format dictated by the applier: SEARCH/REPLACE for Aider diff, unified diff for git, `str_replace` for Anthropic editor tool. | ### Step 3 — Which model am I targeting? Format support is **model-specific**. Aider maintains a per-model edit-format default precisely because of this. [aider.chat/docs/more/edit-formats.html] | Model family | Best code-edit format | Notes | |---|---|---| | GPT-4 Turbo / GPT-4o | `udiff` or SEARCH/REPLACE diff | 20%→61% with udiff on refactor. [aider.chat/2023/12/21/unified-diffs.html] | | Claude 3.5/3.7 Sonnet | SEARCH/REPLACE diff | Tendency to write *too much* — instruct minimal blocks. [aider.chat/2024/07/01/sonnet-not-lazy.html] | | Gemini family | `diff-fenced` | Path inside the fence. [aider.chat/docs/more/edit-formats.html] | | GPT-4.1 / OpenAI patch tool | `patch` protocol | OpenAI-specific, multi-action robust. | | Weak / local models (Llama-3-8B class, GPT-3.5) | `whole` file rewrite | Diff parsing failures dominate; whole-file is dumb-but-stable. | | Reasoning models (o1, o3) as architect | architect mode: reasoner emits prose plan → editor model emits diff | o1-preview alone: 79.7%; o1-preview + Sonnet editor: 82.7%. [aider.chat/2024/09/26/architect.html] | Combined decision tree: ``` What's being emitted? │ ┌───────────────────┼───────────────────┐ │ │ │ code structured free-form edits data fields prose/reasoning │ │ │ ▼ ▼ ▼ diff/SEARCH- JSON / Pydantic / markdown REPLACE/ tool_use scalar str_replace, args code as string │ │ │ ▼ ▼ ▼ pick per-model wrap any code- do NOT edit format shaped payload in force into (Aider table) a single string field JSON schema ``` --- ## 4. 操作模型 (Task → format pairs) Ten concrete operations. Format choice is *not* arbitrary — each citation is the empirical anchor. | # | Task | Recommended format | Rationale / Evidence | |---|---|---|---| | 1 | "Edit `auth.py` to use JWT" (coder-agent core loop) | Plain-text SEARCH/REPLACE diff in markdown fence | Aider's measured 3× on GPT-4 Turbo, generalizes across models. [aider.chat/2024/08/14/code-in-json.html] | | 2 | "Apply this patch to a file via Claude's text-editor tool" | Anthropic `str_replace` tool — old_str / new_str as **string fields**, no JSON sub-structure inside the code | Matches Anthropic's published tool-design guidance: fields should be high-signal scalars, not low-level identifiers. [docs.anthropic.com text-editor-tool] | | 3 | "Extract invoice fields (vendor, amount, date, line items)" | JSON / Pydantic / DSPy typed `OutputField` | Schema **is** the task. Structured-output benchmarks show high accuracy here. [arxiv.org/html/2505.20139v1] | | 4 | "Classify ticket priority (P0/P1/P2/P3)" | Single-token output or JSON `{"priority": "..."}` | Trivial; structure helps determinism. | | 5 | "Answer a math word problem" | Markdown reasoning + final answer in `\boxed{}` or fenced final line | "Let Me Speak Freely?" — 10–15% degradation when locked into JSON reasoning. [arxiv.org/abs/2408.02442] | | 6 | "Decide which tool to call next" | JSON tool_use with tool name + scalar args | Action selection is *intrinsically* structured. | | 7 | "Generate the SQL query for that tool" | Tool args contain `{"sql": "SELECT ..."}` — SQL as a **raw single string**, no further JSON sub-structure | Same family as code-in-JSON; SQL is code-shaped. | | 8 | "Write a customer-support reply" | Markdown | Native to instruction tuning; JSON-wrap costs nothing useful. | | 9 | "Summarize a paper and return key claims as a list" | Markdown prose + a fenced `claims:` YAML or JSON block at the end (two-zone output) | Best of both: prose flows freely, downstream parser reads the trailing block. | | 10 | "Rewrite a long document AND extract entities" | **Two passes**: pass 1 rewrite (markdown), pass 2 extract (JSON, fed the pass-1 output) | One call cannot serve both contents at peak quality. | --- ## 5. 困境决策案例 (Dilemma cases / worked examples) ### Case A — "I'm writing an extractor for invoice fields. Plain text or JSON?" **Trigger**: New extraction pipeline. Fields = `vendor_name`, `total_amount`, `invoice_date`, `line_items[]`. **Constraints**: - Output is consumed by a downstream Python parser. - Some line items contain free-text descriptions with newlines and quotes. - Pydantic validation is desired. **Decision steps**: 1. The content is **structured data**, not code. JSON tax is low here — fields are short scalars or short strings. 2. Use Pydantic + OpenAI `response_format=json_schema` or DSPy typed `OutputField`s. Field names carry the semantic load (DSPy's "Signatures carry semantic load" principle [dspy.ai/learn/programming/signatures]). 3. For `line_items[].description` containing free-text: still inside JSON — descriptions are *prose data*, not code. Escaping cost is real but bounded. 4. Watch for: extraction-then-rewrite mission creep. If a future spec says "also rewrite the invoice text in cleaner language" → switch to Case B's two-pass approach. **Outcome**: JSON wins. Schema validation catches missing fields cheaply; no measurable degradation expected on this content shape. ### Case B — "Same task but the extractor *also* rewrites the invoice text" **Trigger**: PM adds requirement — "and produce a cleaned-up version of the invoice document body." **Constraints**: - "Cleaned text" is multi-paragraph prose with embedded numbers, possibly tables. - Putting `cleaned_text: "..."` inside the same JSON forces multi-paragraph escaping and competes with the extraction reasoning for attention. **Decision steps**: 1. Recognize this as the **mixed-content trap**. One LM call cannot serve both contents at peak. 2. Two solutions, choose by latency budget: - **Two-pass** (preferred): Call 1 returns `{vendor, amount, date, line_items}` as JSON. Call 2 receives the original doc + extracted JSON, returns markdown rewrite. - **Two-zone single call**: Prompt requests markdown rewrite first, then `===EXTRACTION===` separator, then a fenced JSON block. Parser splits on the separator. Cheaper but more fragile. 3. Do not nest the markdown rewrite as a JSON string field. The model will compress it and lose structure. **Outcome**: Plain text + JSON split, not unified JSON. **Format follows content**, even within one logical task. --- ### Case C — "Tool call decision, but the tool's input is SQL" **Trigger**: Building a data-analyst agent. One tool is `run_query(sql: str)`. **Constraints**: - The agent must decide *which* tool to call (structured choice). - The agent must *write SQL* (code-shaped content). - Provider is OpenAI / Anthropic tool_use. **Decision steps**: 1. The **tool-selection layer** is intrinsically structured — use JSON tool_use. The model picks `run_query` vs `read_file` vs `list_tables`. 2. The **tool body** is code. Per Aider's findings, code in JSON degrades. Mitigations: - Make `sql` a **single top-level string field**. Don't sub-structure it (`{from: ..., where: ...}`). Let the model write idiomatic SQL. - Keep other args as scalars (e.g. `dry_run: bool`). - Expect *some* quality hit vs returning SQL as plain text; budget for retry/repair. 3. If quality is unacceptable: switch the architecture. Generate SQL in a *separate* plain-text LM call (markdown fenced ```sql block), then a thin downstream agent wraps it into the tool call. This is the Aider "architect" pattern transplanted to data tools. [aider.chat/2024/09/26/architect.html] **Outcome**: JSON for the decision; single string for the SQL body; consider a two-step generate-then-dispatch if quality matters. --- ### Case D — "Reviewer says 'use structured output everywhere for safety'" **Trigger**: Eng-org-wide push to add `response_format=json_schema` to every LM call. **Constraints**: - The org has been bitten by `json.loads` failures. - Some calls produce code; some produce prose; some produce typed objects. - The reviewer is conflating "schema-valid" with "quality." **Decision steps**: 1. Surface the empirical finding: **JSON validity ≠ content quality.** Cite Aider's 20%→61% (unified diff vs JSON) [aider.chat/2024/08/14/code-in-json.html] and "Let Me Speak Freely?" [arxiv.org/abs/2408.02442]. 2. Audit each call by content type using §3 Step 1 table. 3. For code-emitting calls: keep markdown; add a defensive parser (regex extract from fenced block). The robustness cost is small; the quality gain is large. 4. For typed-object calls: yes, add `json_schema`. This is where structured output earns its keep. 5. Document the policy: **"Structured output is mandatory for typed-data extraction; forbidden for code edits and multi-step reasoning; optional for everything else."** **Outcome**: Reject the blanket policy; replace it with a content-type-driven one. This is the inverse of the "JSON everywhere" reflex. --- ## 6. 反模式与边界 (Anti-patterns & boundaries) ### Anti-patterns 1. **"JSON everywhere" reflex.** Structured ≠ better. Code, multi-step reasoning, long-form prose all degrade inside JSON envelopes — Aider measured 3× on code; arXiv 2408.02442 measured 10–15% on reasoning. Use JSON where the *content itself is structured*.