markform 0.1.7 → 0.1.8

package/README.md

@@ -1,49 +1,66 @@
 # Markform
 
-**Markform** turns Markdown documents into executable specifications for structured data
-collection.
-Define fields, validation rules, and instructions in a single `.form.md` file
-that is readable by humans, parseable by machines, and fillable by LLM agents.
-
-The core idea: **a form combines structure, unstructured context, and memory in a simple
-text document**. Users can give inputs via UIs or CLI. Agents can fill fields
-incrementally via tools or APIs.
-Validation catches errors early and humans can review or intervene at any point.
-The entire workflow is visible in a token-friendly text file you can read, diff, and
-version control.
+**Markform** is a text format for defining structured forms that humans can read,
+machines can parse, and agents can fill via tool calls.
+
+Define instructions, fields, and validation rules in a single `.form.md` file.
+Agents fill forms incrementally via patches.
+Fields are validated, so errors are caught early and can be corrected.
+Humans can review or intervene at any point.
+
+**Why forms?** For deep research or complex AI tasks, you need more than just prompts or
+flow: you need *structure*, which is precise control over agent output at every stage of
+a workflow. A well-designed form combines instructions, structured data, and validations
+in one place.
+
+**How it works:**
 
-Syntax is [Markdoc](https://github.com/markdoc/markdoc), which is Markdown extended with
-`{% tag %}` annotations, so LLMs are already quite good at writing Markform docs.
+- A Markform document exposes a programmatic interface: users fill fields via CLI or web
+  UI, agents fill via tool calls ([Vercel AI SDK](https://github.com/vercel/ai)
+  integration included).
+
+- Changes are explicit patch operations (`{ "op": "set_string", "fieldId": "name",
+  "value": "Alice" }`) validated against a schema specified in the form.
+  The agent sees validation errors and can self-correct.
+
+- The format extends Markdown with a
+  [precise specification](https://github.com/jlevy/markform/blob/main/docs/markform-spec.md).
+  Export Markform syntax to JSON, YAML, JSON Schema, or plain Markdown reports.
+
+Markform syntax is a good source format: token-efficient text you can read, diff, and
+version control.
+It is built with [Markdoc](https://github.com/markdoc/markdoc), which is
+a Markdown extension from Stripe with Jinja-style `{% tag %}` annotations.
 
-## Why?
+## Why Do Agents Need Forms?
 
-Many agent workflow frameworks emphasize *prompts* and the *flow* of information (the
-*how*) over the desired *structure* of the results (the *what*). Markform lets you build
-agent workflows by structuring and validating *what* you want (the structure of
-information, validations, and reviews encoded in a form) instead of *how* to run a
-workflow as code (via explicit workflows or just an unstructured swarm of agents).
+For centuries, humans have used paper forms and
+[checklists](https://en.wikipedia.org/wiki/The_Checklist_Manifesto) to systematize
+complex processes. A form with instructions, field definitions, and validations is a
+concise way to share context: goals, background knowledge, process rules, and state
+(memory). I don’t think AI changes this essential aspect of knowledge work.
 
-For centuries, humans have used paper forms to systematize and manage processes.
-A well-designed form with instructions, field definitions, and validations is a concise
-way to share context: background knowledge, goals, process rules, and memories.
-I don’t think AI changes this essential aspect of knowledge work.
-It’s time to bring bureaucracy to the agents.
+Most agent frameworks focus on *prompts* and *flow* (the how) over the *structure* of
+results (the what).
+But for deep research or other multi-step workflows, you need precise
+control over intermediate states and final output.
+You don’t want that structure in a GUI (not token-friendly), in code (hard to update),
+or dependent on model whims (changes unpredictably with model updates).
 
-There’s one more key benefit to this approach: LLMs are good at writing forms!
-Because the syntax is just Markdown with Jinja-style tags, agents can convert an
-informal Markdown doc describing a process to a precise Markform process easily.
+Forms solve this. Forms codify operational excellence.
+They’re easy to read, easy to edit, and enforce standards.
+Because LLMs handle Markdown and Jinja-style tags well, agents can also help create and
+improve the forms themselves—closing the meta-loop.
 
-(For more, see [the FAQ](#faq).)
+It’s time to bring bureaucracy to the agents!
+See [the FAQ](#faq) for more on the design.
 
 ## Quick Start
 
 ```bash
-# Copy example forms to ./forms/ and run one interactively
+# Copy example forms to ./forms/ and run one interactively.
 # Set OPENAI_API_KEY or ANTHROPIC_API_KEY (or put in .env) for research examples
-npx markform examples  # Copy examples to ./forms/, then offers to run one
-
-# Or run forms directly
-npx markform run  # Browse and run forms interactively
+npx markform@latest examples
 
 # Read the docs (tell your agents to run these; they are agent-friendly!)
 npx markform  # CLI help
@@ -52,7 +69,8 @@ npx markform docs  # Quick reference for writing Markforms
 npx markform spec  # Read the full spec
 ```
 
-The `markform examples` command copies sample forms locally and prompts you to run one.
+The `markform examples` command copies some sample forms to `./forms` and prompts you to
+fill in a form interactively and then optionally have an agent complete it.
 Pick `movie-research-demo.form.md` for a quick example.
 
 ## Installation
@@ -72,56 +90,92 @@ npm install markform
 ### Form Definition
 
 A `.form.md` file is simply a Markdoc file.
-It combines YAML frontmatter with Markdoc-tagged content:
+It combines YAML frontmatter with Markdoc-tagged content.
 
-```markdown
+The text can be any Markdown.
+The tags define things like fields:
+
+```jinja
+{% field kind="string" id="movie" label="Movie" role="user"
+   required=true minLength=1 maxLength=300 %}{% /field %}
+
+{% field kind="single_select" id="mpaa_rating" role="agent" label="MPAA Rating" %}
+- [ ] G {% #g %}
+- [ ] PG {% #pg %}
+- [ ] PG-13 {% #pg_13 %}
+- [ ] R {% #r %}
+- [ ] NC-17 {% #nc_17 %}
+- [ ] NR/Unrated {% #nr %}
+{% /field %}
+```
+
+Fields have types defined by the attributes.
+Values are filled in incrementally, just like any form.
+Once filled in, values appear directly inside the tags, in Markdown format:
+
+```jinja
+{% field kind="string" id="movie" label="Movie" role="user"
+   required=true minLength=1 maxLength=300 %}
+The Shawshank Redemption
+{% /field %}
+
+{% field kind="single_select" id="mpaa_rating" role="agent" label="MPAA Rating" %}
+- [ ] G {% #g %}
+- [ ] PG {% #pg %}
+- [ ] PG-13 {% #pg_13 %}
+- [X] R {% #r %}
+- [ ] NC-17 {% #nc_17 %}
+- [ ] NR/Unrated {% #nr %}
+{% /field %}
+```
+
+Note fields can have a `role="user"` to indicate they are filled interactively by the
+user, or a `role="agent"` to indicate an agent should fill them in.
+
+There are also tags for user or agent instructions per field or at form level and
+grouping of forms.
+
+Checkboxes and tables as values are supported!
+Checkboxes can be single-select or multi-select.
+
+Here’s a full example:
+
+<details>
+
+<summary>Markform for Movie Research Demo (click to expand)</summary>
+
+```jinja
 ---
 markform:
   spec: MF/0.1
-  title: Movie Research (Demo)
-  description: Quick movie lookup with just the essentials (title, year, ratings, summary).
+  title: Movie Research Demo
+  description: Movie lookup with ratings from IMDB and Rotten Tomatoes.
+  run_mode: research
   roles:
     - user
     - agent
   role_instructions:
     user: "Enter the movie title."
     agent: |
-      Quickly identify the movie and fill in basic info from IMDB.
-      This is a demo lookup - just get the core facts.
+      Identify the movie with web searches and use imdb.com and rottentomatoes.com to fill in the ratings.
 ---
-{% form id="movie_research_demo" title="Movie Research (Demo)" %}
-
-## Movie Research Example
+{% form id="movie_research_demo" %}
+{% group id="movie_input" %}
 
-{% group id="movie_input" title="Movie Identification" %}
-
-What movie do you want to research? \[*This field is filled in by the user (`role="user"`).*\]
+## What movie do you want to research?
 
 {% field kind="string" id="movie" label="Movie" role="user" required=true minLength=1 maxLength=300 %}{% /field %}
 {% instructions ref="movie" %}Enter the movie title (add year or details for disambiguation).{% /instructions %}
 
 {% /group %}
 
-## About the Movie
-
 {% group id="about_the_movie" title="About the Movie" %}
 
-**Title:**
-
-{% field kind="string" id="full_title" label="Full Title" role="agent" required=true %}{% /field %}
-{% instructions ref="full_title" %}Official title, including subtitle if any.{% /instructions %}
-
-**Release year:**
-
-{% field kind="number" id="year" label="Release Year" role="agent" required=true min=1888 max=2030 %}{% /field %}
-
-**IMDB:**
-
-{% field kind="url" id="imdb_url" label="IMDB URL" role="agent" required=true %}{% /field %}
+## Movie Ratings
 
-**MPAA rating:**
+Here are the ratings for the movie:
 
-{% field kind="single_select" id="mpaa_rating" label="MPAA Rating" role="agent" %}
+{% field kind="single_select" id="mpaa_rating" role="agent" label="MPAA Rating" %}
 - [ ] G {% #g %}
 - [ ] PG {% #pg %}
 - [ ] PG-13 {% #pg_13 %}
@@ -130,56 +184,181 @@ What movie do you want to research? \[*This field is filled in by the user (`rol
 - [ ] NR/Unrated {% #nr %}
 {% /field %}
 
-**IMDB rating:**
-
-{% field kind="number" id="imdb_rating" label="IMDB Rating" role="agent" min=1.0 max=10.0 %}{% /field %}
-{% instructions ref="imdb_rating" %}IMDB user rating (1.0-10.0 scale).{% /instructions %}
-
-**Summary:**
+{% field kind="table" id="ratings_table" role="agent"
+   label="Ratings" required=true
+   columnIds=["source", "score", "votes"] columnTypes=["string", "number", "number"]
+   minRows=0 maxRows=3 %}
+| Source | Score | Votes |
+|--------|-------|-------|
+{% /field %}
 
-{% field kind="string" id="logline" label="One-Line Summary" role="agent" maxLength=300 %}{% /field %}
-{% instructions ref="logline" %}Brief plot summary in 1-2 sentences, no spoilers.{% /instructions %}
+{% instructions ref="ratings_table" %}
+Fill in scores and vote counts from each source:
+- IMDB: Rating (1.0-10.0 scale), vote count
+- RT Critics: Tomatometer (0-100%), review count
+- RT Audience: Audience Score (0-100%), rating count
+{% /instructions %}
 
 {% /group %}
-
 {% /form %}
 ```
 
+</details>
+
 ### Form Report Output
 
 Run `npx markform examples` to copy examples, then `npx markform run` and select `Movie
-Research (Demo)` to fill it.
-The report output looks like:
+Research Demo` to fill it.
 
-```markdown
-# Movie Research (Demo)
+A form can be exported
+
+- as the filled form (Markform format, just like the input)
+
+- as a report (plain Markdown)
 
-## Movie Identification
+- as values (YAML or JSON)
 
+- as a JSON schema (just the structure)
+
+The report output (using `gpt-5-mini` to fill it in) looks like:
+
+```markdown
 Movie:
-shawshank redemption
+The Shawshank Redemption
 
 ## About the Movie
 
-Full Title:
+MPAA Rating:
+R
+
+Ratings:
+| Source | Score | Votes |
+| --- | --- | --- |
+| IMDB | 9.3 | 3100000 |
+| RT Critics | 89 | 146 |
+| RT Audience | 98 | 250000 |
+```
+
+Here is the schema and YAML values for the form above.
+
+<details> <summary>Filled Markform (click to expand)</summary>
+
+````jinja
+{% form id="movie_research_demo" %}
+
+{% group id="movie_input" %}
+
+{% field kind="string" id="movie" role="user" label="Movie" maxLength=300 minLength=1 required=true %}
+```value
 The Shawshank Redemption
+```
+{% /field %}
 
-Release Year:
-1994
+{% instructions ref="movie" %}
+Enter the movie title (add year or details for disambiguation).
+{% /instructions %}
 
-IMDB URL:
-https://www.imdb.com/title/tt0111161/
+{% /group %}
 
-MPAA Rating:
-R
+{% group id="about_the_movie" title="About the Movie" %}
+
+{% field kind="single_select" id="mpaa_rating" label="MPAA Rating" %}
+- [ ] G {% #g %}
+- [ ] PG {% #pg %}
+- [ ] PG-13 {% #pg_13 %}
+- [x] R {% #r %}
+- [ ] NC-17 {% #nc_17 %}
+- [ ] NR/Unrated {% #nr %}
+{% /field %}
+
+{% field kind="table" id="ratings_table"
+  columnIds=["source", "score", "votes"] columnLabels=["Source", "Score", "Votes"]
+  columnTypes=["string", "number", "number"]
+  label="Ratings" maxRows=3 minRows=0 required=true %}
+| Source | Score | Votes |
+| --- | --- | --- |
+| IMDB | 9.3 | 3100000 |
+| RT Critics | 89 | 146 |
+| RT Audience | 98 | 250000 |
+{% /field %}
+
+{% instructions ref="ratings_table" %}
+Fill in scores and vote counts from each source:IMDB: Rating (1.0-10.0 scale), vote countRT Critics: Tomatometer (0-100%), review countRT Audience: Audience Score (0-100%), rating count
+{% /instructions %}
 
-IMDB Rating:
-9.3
+{% /group %}
+
+{% /form %}
+````
+
+</details>
+
+<details> <summary>YAML Export (click to expand)</summary>
+
+```yaml
+values:
+  movie:
+    state: answered
+    value: The Shawshank Redemption
+  mpaa_rating:
+    state: answered
+    value: r
+  ratings_table:
+    state: answered
+    value:
+      - source: IMDB
+        score: 9.3
+        votes: 3100000
+      - source: RT Critics
+        score: 89
+        votes: 146
+      - source: RT Audience
+        score: 98
+        votes: 250000
+```
 
-One-Line Summary:
-Convicted banker Andy Dufresne is sent to Shawshank State Penitentiary, where he forms an unexpected friendship with inmate Red while holding onto hope and striving to maintain his dignity in a corrupt prison system.
+</details>
+
+<details> <summary>JSON Schema (click to expand)</summary>
+
+```json
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "movie_research_demo",
+  "type": "object",
+  "properties": {
+    "movie": {
+      "type": "string",
+      "title": "Movie",
+      "minLength": 1,
+      "maxLength": 300
+    },
+    "mpaa_rating": {
+      "type": "string",
+      "enum": ["g", "pg", "pg_13", "r", "nc_17", "nr"],
+      "title": "MPAA Rating"
+    },
+    "ratings_table": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "properties": {
+          "source": { "title": "Source", "type": "string" },
+          "score": { "title": "Score", "type": "number" },
+          "votes": { "title": "Votes", "type": "number" }
+        }
+      },
+      "title": "Ratings",
+      "minItems": 0,
+      "maxItems": 3
+    }
+  },
+  "required": ["movie", "ratings_table"]
+}
 ```
 
+</details>
+
 ### More Example Forms
 
 The package includes example forms.
@@ -200,6 +379,72 @@ View them with `markform examples --list`, copy with `markform examples`, and ru
 
 - [`earnings-analysis.form.md`](https://github.com/jlevy/markform/blob/main/packages/markform/examples/earnings-analysis/earnings-analysis.form.md)
   \- Financial analysis form.
+## Architecture
+
+This repo has a specification and an implementation.
+The implementation is a TypeScript API with Vercel AI SDK integration, and a CLI
+interface.
+
+```mermaid
+flowchart LR
+    subgraph SPEC["<b>MARKFORM SPEC</b>"]
+        direction TB
+
+        subgraph L1["<b>LAYER 1: SYNTAX</b><br/>Markdoc tag syntax and frontmatter (form, group, string-field, checkboxes, etc.)"]
+        end
+
+        subgraph L2["<b>LAYER 2: FORM DATA MODEL</b><br/>Schema definitions for forms, fields, values (in Zod but mappable to JSON Schema or Pydantic)"]
+        end
+
+        subgraph L3["<b>LAYER 3: VALIDATION & FORM FILLING</b><br/>Rules for filling forms via patches, field ids, required field semantics, validation hooks"]
+        end
+
+        subgraph L4["<b>LAYER 4: TOOL API & INTERFACES</b><br/>Abstract API for agents and humans (TypeScript and AI SDK integration)"]
+        end
+
+        L4 --> L3 --> L2 --> L1
+    end
+
+    subgraph IMPL["<b>THIS IMPLEMENTATION</b>"]
+        direction TB
+
+        subgraph CLI["<b>`markform` CLI</b><br/>Command-line interface to all features"]
+        end
+
+        subgraph AGENT["<b>AGENT TOOL INTERFACE</b><br/>Tool API library using AI SDK tools"]
+        end
+
+        subgraph HARNESS["<b>EXECUTION HARNESS</b><br/>Step-by-step form-filling agentic loop"]
+        end
+
+        subgraph ENGINE["<b>CORE TYPESCRIPT APIS</b><br/>Markdoc parser, serializer, patch application, validation (uses jiti for TypeScript rules)"]
+        end
+
+        subgraph TEST["<b>TESTING FRAMEWORK</b><br/>Golden session testing with .session.yaml transcripts"]
+        end
+
+        CLI --> ENGINE
+        CLI --> HARNESS
+        AGENT --> HARNESS
+        AGENT --> ENGINE
+        HARNESS --> ENGINE
+        ENGINE --> TEST
+    end
+
+    SPEC ~~~ IMPL
+
+    style SPEC fill:#e8f4f8,stroke:#0077b6
+    style L1 fill:#caf0f8,stroke:#0077b6
+    style L2 fill:#caf0f8,stroke:#0077b6
+    style L3 fill:#caf0f8,stroke:#0077b6
+    style L4 fill:#caf0f8,stroke:#0077b6
+    style IMPL fill:#fff3e6,stroke:#fb8500
+    style ENGINE fill:#ffe8cc,stroke:#fb8500
+    style CLI fill:#ffe8cc,stroke:#fb8500
+    style AGENT fill:#ffe8cc,stroke:#fb8500
+    style HARNESS fill:#ffe8cc,stroke:#fb8500
+    style TEST fill:#ffe8cc,stroke:#fb8500
+```
 
 ## CLI Commands
 
@@ -272,11 +517,24 @@ markform export my-form.form.md --format=yaml
 markform dump my-form.form.md
 ```
 
+### Export JSON Schema
+
+```bash
+# Export form structure as JSON Schema (for validation, code generation, etc.)
+markform schema my-form.form.md
+
+# Pure JSON Schema without Markform extensions
+markform schema my-form.form.md --pure
+
+# Specify JSON Schema draft version
+markform schema my-form.form.md --draft draft-07
+```
+
 ### Apply Patches
 
 ```bash
 # Apply a JSON patch to update field values
-markform apply my-form.form.md --patch '[{"op":"set","fieldId":"name","value":"Alice"}]'
+markform apply my-form.form.md --patch '[{"op":"set_string","fieldId":"name","value":"Alice"}]'
 ```
 
 ### Web Interface
@@ -308,88 +566,25 @@ markform models
 markform --help
 ```
 
-## Supported Providers
-
-Standard LLMs can be used to fill in forms or create research reports from form
-templates. The package currently has support for these models built in, and enables web
-search tools for them if possible.
-
-| Provider | Env Variable | Example Models |
-| --- | --- | --- |
-| openai | `OPENAI_API_KEY` | gpt-5-mini, gpt-5.1, gpt-5.2 |
-| anthropic | `ANTHROPIC_API_KEY` | claude-sonnet-4-5, claude-opus-4-5 |
-| google | `GOOGLE_API_KEY` | gemini-2.5-pro, gemini-2.5-flash |
-| xai | `XAI_API_KEY` | grok-4, grok-4-fast |
-| deepseek | `DEEPSEEK_API_KEY` | deepseek-chat, deepseek-reasoner |
+## API Key Setup
 
 Set the appropriate environment variable for your provider before running `markform
-fill`. See
-[`src/settings.ts`](https://github.com/jlevy/markform/blob/main/packages/markform/src/settings.ts)
-for the full list of models.
-
-## Architecture
-
-```mermaid
-flowchart LR
-    subgraph SPEC["<b>MARKFORM SPEC</b>"]
-        direction TB
-
-        subgraph L1["<b>LAYER 1: SYNTAX</b><br/>Markdoc tag syntax and frontmatter (form, group, string-field, checkboxes, etc.)"]
-        end
-
-        subgraph L2["<b>LAYER 2: FORM DATA MODEL</b><br/>Schema definitions for forms, fields, values (in Zod but mappable to JSON Schema or Pydantic)"]
-        end
-
-        subgraph L3["<b>LAYER 3: VALIDATION & FORM FILLING</b><br/>Rules for filling forms via patches, field ids, required field semantics, validation hooks"]
-        end
-
-        subgraph L4["<b>LAYER 4: TOOL API & INTERFACES</b><br/>Abstract API for agents and humans (TypeScript and AI SDK integration)"]
-        end
-
-        L4 -->
-
-L3 --> L2 --> L1
-    end
-
-    subgraph IMPL["<b>THIS IMPLEMENTATION</b>"]
-        direction TB
-
-        subgraph ENGINE["<b>ENGINE IMPLEMENTATION</b><br/>Markdoc parser, serializer, patch application, validation (uses jiti for TypeScript rules)"]
-        end
+fill`:
 
-        subgraph UI["<b>USER INTERFACES</b><br/>CLI commands, web UI (serve), render to HTML"]
-        end
-
-        subgraph AGENT["<b>AGENT INTERFACES</b><br/>Tool API library, MCP server, AI SDK tools"]
-        end
-
-        subgraph HARNESS["<b>EXECUTION HARNESS</b><br/>Step-by-step form-filling agentic loop"]
-        end
-
-        subgraph TEST["<b>TESTING FRAMEWORK</b><br/>Golden session testing with .session.yaml transcripts"]
-        end
-
-        UI --> ENGINE
-        AGENT --> HARNESS
-        AGENT --> ENGINE
-        HARNESS --> ENGINE
-        ENGINE --> TEST
-    end
+| Provider | Env Variable | Native Web Search |
+| --- | --- | :---: |
+| openai | `OPENAI_API_KEY` | ✓ |
+| anthropic | `ANTHROPIC_API_KEY` | ✓ |
+| google | `GOOGLE_API_KEY` | ✓ |
+| xai | `XAI_API_KEY` | ✓ |
+| deepseek | `DEEPSEEK_API_KEY` | ✗ |
 
-    SPEC ~~~ IMPL
+Run `markform models` to see available models.
+See
+[`src/settings.ts`](https://github.com/jlevy/markform/blob/main/packages/markform/src/settings.ts)
+for defaults.
 
-    style SPEC fill:#e8f4f8,stroke:#0077b6
-    style L1 fill:#caf0f8,stroke:#0077b6
-    style L2 fill:#caf0f8,stroke:#0077b6
-    style L3 fill:#caf0f8,stroke:#0077b6
-    style L4 fill:#caf0f8,stroke:#0077b6
-    style IMPL fill:#fff3e6,stroke:#fb8500
-    style ENGINE fill:#ffe8cc,stroke:#fb8500
-    style UI fill:#ffe8cc,stroke:#fb8500
-    style AGENT fill:#ffe8cc,stroke:#fb8500
-    style HARNESS fill:#ffe8cc,stroke:#fb8500
-    style TEST fill:#ffe8cc,stroke:#fb8500
-```
+If unsure, try `gpt-5-mini` first as it’s fast and supports web search.
 
 ## Programmatic Usage
 
@@ -422,10 +617,14 @@ import { createMarkformTools, MarkformSessionStore } from "markform/ai-sdk";
 import { generateText } from "ai";
 import { anthropic } from "@ai-sdk/anthropic";
 
+// Parse form and create session store (tracks state across tool calls)
 const form = parseForm(markdownContent);
 const store = new MarkformSessionStore(form);
+
+// Create tools the agent can call: inspect, apply patches, export
 const tools = createMarkformTools({ sessionStore: store });
 
+// Agent fills the form via tool calls until complete or maxSteps reached
 const result = await generateText({
   model: anthropic("claude-sonnet-4-5-20250929"),
   prompt: "Fill out this form with appropriate values...",
@@ -443,11 +642,32 @@ const result = await generateText({
 | `markform_export` | Export schema and values as JSON |
 | `markform_get_markdown` | Get canonical Markdown representation |
 
+## Other Documentation
+
+- **[Quick
+  Reference](https://github.com/jlevy/markform/blob/main/docs/markform-reference.md)**
+  (or run `markform docs`) — Concise syntax reference (agent-friendly)
+
+- **[Markform Spec](https://github.com/jlevy/markform/blob/main/docs/markform-spec.md)**
+  (or run `markform spec`) — Complete syntax and semantics
+
+- **[API
+  Documentation](https://github.com/jlevy/markform/blob/main/docs/markform-apis.md)**
+  (or run `markform apis`) — TypeScript and AI SDK APIs
+
+- **[Design
+  Doc](https://github.com/jlevy/markform/blob/main/docs/project/architecture/current/arch-markform-design.md)**
+  — Technical design and roadmap
+
+- **[Development](https://github.com/jlevy/markform/blob/main/docs/development.md)** —
+  Build, test, and contribute
+
 ## FAQ
 
-### Can you say more why this is a good idea?
+### Can you talk more about why forms are cool?
 
-Yes! I’ve come to believe forms are a missing piece of the workflow problem with agents.
+As a matter of fact, I can!
+I’ve come to believe forms are a missing piece of the workflow problem with agents.
 For deep research or complex multi-step workflows, key pieces need to be *precisely
 controlled*, *domain-specific*, and *always improving*. You need precise documentation
 on the key intermediate states and final output from an AI pipeline.
@@ -458,6 +678,27 @@ Forms define these pieces and are easy to edit.
 All other choices can be left to the agents themselves, with the structure and
 validations enforced by the form-filling tools the agents use.
 
+Another cool thing about forms: they get rid of the inefficiencies of conversation chat
+history.
+Often when an agentic loop is built, it just saves the chat history for context.
+But a form is inherently more efficient: the harness itself can be stateless.
+It just shares the partly-filled form with the agent, and it has full context in one
+message. That’s what the agentic loop in this implementation does.
+
+Finally, the meta-loop of *creating and improving* forms is easier to automate:
+
+- To get started, you can ask a good coding model to convert any unstructured doc
+  describing a process to a form.
+  The model can also use the CLI or tools to validate and test it.
+
+- Any time you have a workflow problem, you can ask an LLM to diagnose it and if
+  possible, go back and fix up the form with additional instructions or fields or checks
+  that would prevent the problem from happening again.
+
+I suspect dynamic form structures like this could make complex deep research more
+powerful. Just as you plan a spec before implementing with a coding agent, you could use
+Markform to encode a research plan before dispatching agents to fill it.
+
 ### Is this mature?
 
 No! I just wrote it.
@@ -478,11 +719,11 @@ See
 for more thoughts on how this works.
 And see [the complete history of
 specs](https://github.com/jlevy/markform/tree/main/docs/project/specs/done) for examples
-of how everything is done with specs and with
+of how everything is done with specs.
 
-Although I didn’t have to write much there was a *lot* of management and review by me
-and a lot of thought and iteration for all design decisions.
-And this doc is written by me.
+Although I didn’t write much code, there was a *lot* of management, review, and
+iteration on design decisions.
+And yes, this README is written by me.
 :)
 
 ### What are the goals of Markform?
@@ -539,13 +780,14 @@ The closest alternatives are:
   human-friendly UI. But these do not have a human-friendly text format for use by
   agents as well as humans.
 
-| Approach | Has GUI | Human-readable source format | Agent-editable | APIs and validation rules |
+| Approach | Usable GUI editor | Human-readable source format | Agent-editable | APIs and validation rules |
 | --- | :---: | :---: | :---: | :---: |
-| Plain Markdown | ☑️ existing tools | ✅ | ⚠️ fragile | ❌ |
-| JSON with schema | ⚠️ in some apps | ⚠️ um it’s JSON | ✅ | ✅ |
-| SaaS tools (Typeform, Docusign, PDF forms) | ✅ | ⚠️ rarely | ⚠️ if they add it | ⚠️ if they add it |
-| Excel/Google Sheets | ✅ | ❌ .csv (poor) or .xlsx (worse) | ⚠️ with tools | ✅ with some coding |
-| **Markform** | ☑️ existing tools | ✅ | ✅ with this package | ✅ with this package |
+| Plain Markdown | ✅ IDEs/editors | ✅ | ⚠️ fragile | ❌ |
+| JSON + JSON Schema | ✅ IDEs/editors | ⚠️ no free text | ✅ | ✅ |
+| SaaS tools (Typeform, Docusign, PDF forms) | ✅ | ⚠️ rarely | ⚠️ sometimes | ⚠️ sometimes |
+| HTML/web Forms | ✅ IDEs/editors | ⚠️ HTML+code | ⚠️ coding agent | ✅ |
+| Excel/Google Sheets | ✅ app | ❌ .csv/.xlsx | ⚠️ with tools | ✅ with some coding |
+| **Markform** | ✅ IDEs/editors | ✅ | ✅ with this package | ✅ with this package |
 
 ### What are example use cases?
 
@@ -587,26 +829,6 @@ settings:
 
 Or see [markdoc/language-server](https://github.com/markdoc/language-server).
 
-## Documentation
-
-- **[Quick
-  Reference](https://github.com/jlevy/markform/blob/main/docs/markform-reference.md)**
-  (or run `markform docs`) - Concise syntax reference (agent-friendly)
-
-- **[Markform Spec](https://github.com/jlevy/markform/blob/main/docs/markform-spec.md)**
-  (or run `markform spec`) - Complete syntax and semantics
-
-- **[API
-  Documentation](https://github.com/jlevy/markform/blob/main/docs/markform-apis.md)**
-  (or run `markform apis`) - TypeScript and AI SDK APIs
-
-- **[Design
-  Doc](https://github.com/jlevy/markform/blob/main/docs/project/architecture/current/arch-markform-design.md)**
-  \- Technical design and roadmap
-
-- **[Development](https://github.com/jlevy/markform/blob/main/docs/development.md)** -
-  Build, test, and contribute
-
 ## License
 
 AGPL-3.0-or-later. [Contact me](https://github.com/jlevy) for additional licensing