Skip to content
DOCS

DOCS / READING OUTPUTS

Reading outputs

A stored completion is one view of a call. Four other formats exist for reading it again. This page records how old each one is, and which two the dashboard builds.

Four formats

As models write more of the first draft, more of the work is reading what they wrote. People already ask a model to restate an answer in a tighter form. The ladder is the same each time: shorter sentences, then a picture, then a page, then a video.

  1. 01 Controlled writing, in the style of ASD-STE100.
  2. 02 A diagram.
  3. 03 A self-contained HTML page.
  4. 04 A narrated explainer video.

None of the four is new. The new habit is to spend the model's own output on oversight, and to throw the artifact away when the reading is done.

Controlled writing

ASD-STE100 is Simplified Technical English. Aerospace maintenance writers started it in 1983 as AECMA Simplified English, so a technician would not misread a procedure. The guide followed in 1986. The name ASD-STE100 dates to 2005. Issue 9, released on 15 January 2025, is an international standard. The owner is ASD, the AeroSpace and Defence Industries Association of Europe. The first field was aircraft maintenance. Defence, rail, and other industries use it now.

The specification has two parts. Part 1 is a set of writing rules, nine sections, about 53 rules. Part 2 is a dictionary of about 900 approved words. Each approved word has one meaning and one part of speech. The full text is a free download from asd-ste100.org. This repo does not include that dictionary.

Limits people remember from the writing rules:

  • 20 words in a procedure sentence. 25 words in a description sentence.
  • 6 sentences in a descriptive paragraph. One topic in that paragraph.
  • 3 words in a noun cluster.
  • Procedures use the imperative, the simple present, the simple past, and the infinitive.
  • Procedures avoid the -ing form, the perfect, and the passive.
  • One instruction per sentence. Put the condition before the command.

Checkers for this style are old. Etteplan's HyperSTE has checked Simplified Technical English since about 2002, inside authoring tools and later as a cloud service. The STEMG white paper on AI says models may draft and translate this language, and that the standard still wins when the two disagree. ASD certifies no checker. A prompt can imitate the style. It cannot guarantee the dictionary.

Using the style as a way to read model output, rather than as a way to write a maintenance manual, showed up in agent skills in 2025 and 2026. "80% of the way to ASD-STE100" is the honest target. Full conformance is the wrong claim for a prompt.

Diagrams

Asking a model for Mermaid, SVG, or a flowchart has been ordinary since 2023. Chat products render those diagrams in the thread. Observability tools already draw the same idea from a trace: Langfuse builds an agent graph from the steps, both collapsed by name and expanded in run order.

A diagram of a call you already stored does not need another model. The participants and the order are in the log.

Web pages

Claude Artifacts shipped in June 2024. A side panel renders HTML, React, SVG, and Mermaid as their own object. ChatGPT Canvas shipped in October 2024 as a document and code surface beside the chat. Later previews render HTML in the thread as well. Products such as v0, Bolt, and Lovable are built around "ask for a page."

An HTML explainer of a trace is possible later, inside a sandbox. Model HTML is untrusted script. It needs a locked frame, no dashboard cookies, and a size cap. This version does not build that sandbox.

Explainer videos

The 3Blue1Brown look comes from Manim. The community edition has been public since about 2020. Narrated explainers of your own sources are a product already: NotebookLM Audio Overviews shipped on 11 September 2024, and Video Overviews shipped on 29 July 2025 as narrated slides.

A fully custom animation is the step that is only starting to hold. Open pipelines such as STEMViz, Chalkboard, and manim-agent write Manim code, render it, add speech from ElevenLabs or a local voice, and retry when the render fails. They demo well. They are not a reliable button inside a cost dashboard.

A video also needs a render worker. It does not run in this web app. A short video costs more time and money than the call it explains. This version does not render video.

In the dashboard

Open a call, or open a trace. Three views sit on the stored text:

  • Plain. The messages and the waterfall you already had.
  • STE. A local reading aid. It shortens a few phrases and lists notes for sentence length, passive voice, and -ing forms.
  • Diagram. A sequence of the call or the trace, plus Mermaid you can copy.

The STE view and the diagram run in the browser on text the gateway already stored. The preview cost is $0.00. They do not add a gateway call. A later version can send the same rewrite through the gateway and put that call on the receipt.

Open Calls and choose a row, or open Traces and choose a run.

Limits

The STE tab is a reading aid. It is not an ASD-STE100 check. It does not load the official dictionary. A sentence can pass this tab and still fail the specification.

Word swaps, and only these:

  • in order to → to
  • prior to → before
  • in the event that → if
  • utilize → use
  • approximately → about
  • commence → start
  • ensure → make sure

Inflected forms follow the same swap. "Utilized" becomes "used". "Ensures that" becomes "makes sure".

Notes flag a procedure sentence over 20 words and a description sentence over 25 words. A sentence counts as a procedure when its first word is not a subject or an article. Notes flag a passive such as "was closed". Notes flag a word of five letters or more that ends in "ing", except during, nothing, something, anything, and everything. The checker does not know the part of speech, so a noun such as "billing" can raise a note. Notes do not delete a fact to force the word limit.

HTML pages and explainer videos stay documented here and unimplemented in the app.