Wright

Drop in the requirement files, say one sentence, and the page appears. Change it by talking. When it is right, take it all the way to SAP.

Everything below was done today on a Mac with the installed app, in English. The example is the Sysmex R-043 Sales Details Report, built from its 4 original requirement files: the FSD/TSD, two field-mapping sheets and one sample-data sheet.

20 sto the first full page

9 sto point at a spot and change it

1 min 10 sto install on SAP test and check it

4AI engines, switchable and compared

01Files in, page out

Paste or drag the requirement files in and say "build a sales details report page". No long brief and no waiting for the whole job: after 20 seconds a first full page is on the right, and it fills in while you watch. The left side always says what it is doing.

Fresh project, welcome screen — A fresh project. Short prompts in English or Chinese are fine.

Four files pasted into the message box — The 4 requirement files, copied in Finder and pasted into the message box.

First full version after 20 seconds — 20 s"Build a sales details report page from these files, with sample data for now." A first full page is already on the right; each file shows "saved in project".

The full report page — 6 min 9 sThe full page: three report tabs, all selection fields, totals, an Amounts switch and the detail table. Before designing, it looked up on ONQ 200 (read-only) which tables and function modules in the spec already exist, and it checked the page at two laptop widths before handing it over.

02Point at it and say what to change

To change something, hold ⌥ and click it on the page, then say "make this row's font a bit bigger". No need to explain which part you mean.

Pointing at the table header — ⌥-click on the table header: it gets an outline, and the message box shows "Pointed at: Sold To Party Name".

Header row font bigger — 9.4 s"Make this row's font a bit bigger." Only the header row grows, on all three tabs; nothing else moves.

03Ask for Fiori, get the Fiori look

One sentence, "make a SAP Fiori style version", and the first version already looks like Fiori: shell bar, filter bar, key-figure cards and a table card in SAP's own colours and fonts. It starts as an HTML page so you can see it; once you are happy it can become a real Fiori app.

Fiori first version — 30 sThe first Fiori version.

Full Fiori page — 1 min 48 sThe full Fiori page: tabs, filter bar with Go and Adapt Filters, key-figure cards.

04All the way to SAP, safely

When the page is right, say "install it on BTQ 100". Three safeguards hold the whole way:

Look before designingBefore designing, it reads ONQ (read-only) to see which tables, function modules and Z objects already exist, and reuses them.

Ask before every writeEach time it wants to write to SAP, a card says which system, which object and what it will do. Nothing happens until you click Allow.

Clean up afterIt writes only in BTQ's temporary area ($TMP), with no transport request. One click deletes everything, and it checks again that it is gone.

Approval card before writing to SAP — 35 s"Claude wants to write on SAP and asks you first": system BTQ 100, the class and the page service, the action in $TMP. Allow or Don't allow.

Installed and checked — 1 min 10 sInstalled on BTQ 100, then checked straight away: syntax check, ATC and authorization. (The 564 ATC notes are all one harmless kind about page text; the newer build counts only real program findings.)

Summary and SICF steps — SAP has no way to create the SICF service node automatically, so Wright lists the five clicks to do it by hand, in plain words.

Service check — 10 sAfter the five clicks: "The page opens from BTQ 100." The service is active and answers. (Server address greyed here.)

The page served by BTQ — The same report as BTQ 100 serves it, opened in the browser with the normal SAP login.

The SAP panel: BTQ may be written in $TMP, ONQ and ONP are read-only, "Ask me before writing" is on, and every object Wright made is listed.

Delete all done — 12 sDelete all: "Deleted the 1 object Wright created, and checked TADIR: it is gone." The SICF node is removed by hand first, as Wright explains. Our own check afterwards: nothing left.

05Working comfortably

Small things that make the day easier: you always see what it is doing, several jobs can run side by side, a screenshot works as an instruction, and Wright speaks your language.

Live status line — 14 sNever silent: the status line says what it is changing right now and for how long (here "changing index.html, 10 lines written · 14 s", on the Chinese screen), with each finished step ticked above.

Two sessions side by side — 22 sTwo sessions at once: one is still replying while the other is done and waiting for you.

Pasted screenshot as instruction — Paste a screenshot and say "add four summary cards above the table, like this".

Cards added in the screenshot's style — 31 sThe four cards appear in the screenshot's style. fixed sinceThe reply came in English after a Chinese prompt; that is fixed in the newer build.

English question answered in English — 5.2 sAsk in English, get the answer in English: three short bullets about the page.

Settings: language and appearance — Settings: Wright's own language (中文 / English) and appearance (system, light, dark). The assistant always answers in the language you write in.

Dark mode. The page being built keeps its own look.

"Open in browser" shows the page full size; later changes appear there too.

Skills that learn, within rules

Your team's know-how lives in a skills library. Every change becomes a new version you can go back to, and a lesson only enters the library if it passes the written rules (the "constitution").

Lessons checked against the rules — ~1 min 40 s"Collect this session's lessons": one lesson found, checked against the rules and dropped, with the reason shown.

Every change is recorded

End session dialog — Each session works on its own copy. When it ends you choose: merge its 9 changed files into the project, or discard them.

Project folder with records — After merging: the pages, the instruction log, the change notes for the session and the run record, all in the project folder and in version history.

06What each step costs; switch engines any time

Under every reply: how long it took and what it cost. On the Claude subscription it shows what the same use would cost at API prices. Within one conversation you can switch engines; the project files and the conversation so far go to the new engine and it carries on.

Cost line under each reply — Under each reply: "Took 1 min 48 s … API equiv. $0.671". Bottom left: the whole project so far.

Price table — The price table is a plain file, with the source and the date each price was read. Edit a number and new turns use it.

07Four engines, one report, side by side

Same 4 requirement files, same 4 sentences, same starting point. The four engines run one after another and are compared on page quality, time, cost and whether a person had to step in. Whatever each engine creates on SAP is deleted at the end.

Today's run on the demo Mac: finished

Started 13:08, all four engines done at 15:31 (2 h 23 min end to end). Every engine answered all 4 sentences with nobody stepping in, every engine asked before each SAP write, and every engine's SAP objects were deleted and checked. Total paid: $9.48, well under the $19 of caps; no engine hit its cap.

Engine	All 4 sentences done?	Time	Cost	Allow clicks for SAP	SAP cleaned up
Claude subscription	Done (4 of 4)	18 min 50 s *	monthly plan; $4.96 at API prices	1	Yes, checked
Claude API	Done (4 of 4)	27 min 48 s *	$6.14	3	Yes, checked
DeepSeek	Done (4 of 4)	16 min 39 s	$0.38	2	Yes, checked
Zhipu GLM	Done (4 of 4)	74 min 18 s *	$2.96	3	Yes, checked

* includes time the run sat waiting for a person to click Allow: about 10 min each for the two Claude engines, about 4 min for GLM. The tester was slow there, not the engine. Our own check on BTQ afterwards: nothing from this run is left.

In one line, today: DeepSeek was the fastest and by far the cheapest; the Claude subscription was close behind at no extra charge; Claude API cost the most; GLM finished this time but took over an hour. Speed and cost are measured; which page reads best is for you to judge from the shots.

Final side-by-side view of the four engines — 2 h 23 minThe final view: the four pages side by side, each engine's figures underneath. All four answered 4 of 4.

SAP row: created, deleted, checked — The full figures down to the SAP row (outlined): every engine created 1 object on BTQ, deleted it and checked it was gone. Nobody stepped in on any engine.

Along the way

Comparison start sheet — One sheet starts it all: the task folder, the four engines with a spending cap each, and the 4 sentences every engine gets.

Claude subscription finished — 18 min 50 sClaude subscription done: all 4 sentences answered, its page shown live in its column, SAP cleaned up.

Claude API asks before writing to SAP — Even in a comparison run, each engine asks before writing to SAP: the same Allow card, here for Claude API.

Claude API finished — 27 min 48 sClaude API done: 4 of 4, $6.14. It fixed a code-check finding on its own and re-installed.

DeepSeek finished — 16 min 39 sDeepSeek done: 4 of 4 for $0.38, about one sixteenth of Claude API's cost. SAP cleaned up.

GLM running — 21 min 54 sGLM midway: 3 of 4 sentences sent. It went on to finish all 4 in 74 min 18 s for $2.96.

Earlier runs on the same report (yesterday and this morning)

Engine	All 4 sentences done?	Time	Cost
Claude subscription	Done	about 13–16 min	monthly plan; about $3.5–5.1 at API prices
Claude API	Done	12.5 min	$3.30
DeepSeek	Done	27 min	$0.35
Zhipu GLM	Not finished (hit its cost cap)	78–84 min	up to $5.10

Earlier runs ranked differently (GLM hit its cap then; DeepSeek was slower). Results vary run to run, which is why Wright lets you choose the engine per project and budget, and compare them on your own report.

08Progress

The first release has 13 parts. 12 are done and checked on the real app; the last one, this dry run and package, is in progress today.

12 of 13

1Wright runs as its own Mac app, one conversation with the assistantdone 11-08
2Several conversations run side by sidedone 11-09
3The page being built shows live beside the conversationdone 11-09
4Files and screenshots go straight indone 11-09
5Every instruction and change is recorded and kept in version historydone 11-09
6Reads SAP; writes only in BTQ's temporary areadone 11-09
7The report's page opens from BTQdone 11-09
8Any of four AI engines, behind one switchdone 11-09
9Every reply shows its time and costdone 11-09
10Skills library with versions and written rulesdone 11-09
11The four engines compared on the same reportdone 11-10
12Installs from a signed package, like any Mac appdone 11-09
13Full dry run on the demo Mac and this package for youin progress

Improvements added on 11-09 and 11-10, for what matters most to you

English or ChineseWright's own words in either; replies follow your language.

Dark modeFollows the Mac, or set light or dark.

First version within a minuteA whole page in about 20 s, then filled in.

Point at the page⌥-click a spot, then say what to change.

Fiori lookAsk for Fiori and the first version already looks like it.

Checks its own pageLooks at two laptop widths and fixes cut-off text before handing over.

Looks in SAP before designingRead-only look-up of what already exists.

Code checks after installSyntax, ATC and authorization, explained in plain words.

SAP idle reviveAn idle SAP connection wakes up again by itself.

Honest labels: "done" means checked on the real app on the demo Mac. Two small issues seen in today's shots (the reply language after an engine switch, and the ATC count) are fixed in the newer build and not yet in the installed package.

09How Wright works

One Mac app. Files go in, you talk, the page changes beside the conversation. The engine behind it can be swapped. SAP is read freely, written only with your OK and only in a temporary area, and cleaned up after. Everything is recorded.

10Next

Your own requirement, from scratch: take the 3 original attachments of the SIP sales order dashboard, build it in Wright from zero, and put it next to what you made earlier in Claude.
From one person to a team: today it is one person on one Mac. Next is the platform you designed: a shared skills library, permissions by person and by system, and review plus transport requests up to QA and production.
Try it and tell us: what matters most to you gets polished first.