Think first (qwen38-think)

Before the first edit, write the two tables below into your answer. Then build. End with the THINK line. Doing the work first and the tables after is a fail of this skill.

1. Assumptions table

Everything the brief leaves open that changes what you would build. Take the assumption; do not ask.

Zero rows is allowed only if you write "none — the brief fixes every choice" and can point to where.

2. Three-flaw check — each one checked, not guessed

When a flaw is found: do not stop, do not ask, do not invent the missing thing. Write one line under the table — "Pushback: the brief says A and B; I build A because …" / "Pushback: inventory/report.py does not exist; I did not create it; the nearest file is …" / "Pushback: not possible as stated because …; I did the possible part: …" — and build the part that stands. The developer reads the report; a question asked instead of the work only stops it. In an interactive session you may ask once the report is delivered; in an unattended run there is nobody to answer, so the report is the whole channel.

2b. Thinking is for decisions, not for code

Decide in thinking — the structure, the file order, the risks — then build with tools. For work that spans files, write the order in one line (index.html → css → js/scene.js → js/main.js), then Write the first file. A file drafted in thinking is written twice (the draft, then the Write) and lost if the turn is cut: one GSQ turn was a 43,564-character thinking block with the whole page and both shaders in it, hit the output limit and wrote nothing; from the next turn, one file per turn after short thinking, all four files were done in 15 minutes (#377). Check what you built by running the check (node --check, the tests, the gate), not by re-reading it in thinking.

3. When stuck (a command fails twice, or the output makes no sense)

Write three lines before the next attempt:

Observed: <the exact last line of the failing output>
Hypothesis: <one cause>
Falsify: <one command whose result would rule it out>

Run the falsifying command. If the cause is a missing tool or package: one line "<tool> is not installed; continuing without it" — never install, never search for an alternative, never spawn an agent (using-qwen38).

Report

THINK: assumptions <k> · flaws checked 3 · found <n> · pushback: <yes: the line | no>

The line is the first line of the final report, before the task's own gate line (CODE GATE, GATE) when there is one — a gate report that starts with "Done." and no THINK: line dropped a correct pushback on 2026-09-05 17:38.

What this does not touch

How to build the thing — the task's own skill and gate. This file only makes the model write down what it would otherwise silently assume, and say what is wrong with a brief instead of working around it.