Developers without an AI-integrated IDE keep reaching for the ChatGPT browser tab, and it is genuinely good at one-shot coding help, as long as you do not ask it to remodel your codebase. This is the no-IDE workflow: how to scope a request, paste the right context, pick the right GPT-5.5 mode, and verify the answer in five minutes instead of finding the bug in production three days later.
TL;DR
- ChatGPT is best for self-contained coding: a regex, a single script, a SQL one-liner, a shell pipeline, or decoding an error message. It is not built to edit files across your repo. For that, use Cursor or Claude Code.
- As of June 2026 the default model is GPT-5.5, with a picker on paid plans for Instant (fast), Thinking (reasoning), and Pro ($200 tier). Use Thinking for anything past a one-liner.
- Paste text, not screenshots, keep it under ~300 lines, and always include the exact error message and the command that produced it.
- For anything involving a file (a CSV, a log, a small dataset), use Code Interpreter so ChatGPT runs the Python itself instead of guessing the output.
- Run the code in your own environment before it touches anything that matters. “Looks right” is the number-one source of pasted-and-broken bugs.
Pick the right GPT-5.5 mode
The model you pick matters more than the prompt. Since GPT-5.5 became the ChatGPT default (around April 23, 2026), paid plans expose a picker; the free tier gets GPT-5.5 with no manual mode selection.
| Mode | Best for | Speed | Where |
|---|---|---|---|
| GPT-5.5 Instant | One-liners, regex, syntax lookups, quick rewrites | Fast | Free + all paid |
| GPT-5.5 Thinking | Logic, debugging, anything past ~20 lines | Slower | Go, Plus, Pro, Business |
| GPT-5.5 Pro | Hard, multi-step problems where you cannot afford a wrong answer | Slowest | Pro ($200) |
Rule of thumb: if the task has any branching logic or you would not trust a junior dev to one-shot it, switch to Thinking. The extra wait buys correctness, and on Plus you get up to 3,000 Thinking messages per week (as of June 2026), so there is no reason to hoard them.
What ChatGPT is good at without an IDE
Self-contained work where the whole problem fits in the chat:
- Scripts you run once: a Python data-munge, a bash glue script, a JS snippet.
- Regex you cannot remember the syntax for, and the test cases to prove it works.
- SQL one-liners and dialect conversions (MySQL to Postgres, and the behavioral gotchas).
- Shell pipelines, especially the macOS-vs-Linux flag differences that bite everyone.
- Reading an error message and explaining what it actually means.
- Explaining a library you have never used before, with a minimal working example.
When to switch tools instead
Reach for a real AI coding tool the moment the task needs to see or edit your repo:
- Multi-file refactors, renames, or anything that spans more than the file in your clipboard.
- Work that needs the AI to read your test suite or run your build.
- Changes the AI should apply to disk directly, not hand you to copy-paste.
For those, use Cursor or Claude Code (which runs Claude Opus 4.7 / Sonnet 4.6). See Cursor vs Claude Code to decide between them. ChatGPT in the browser has no project view, so every file you do not paste is invisible to it.
A real session, start to finish
The pattern that works is goal, then code, then constraint, then error:
- State the goal in one sentence, then paste the relevant code. Keep it under ~300 lines. On Plus, the in-app context fits roughly 320 pages of text, but past 300 lines the model spreads its attention thin and the answer degrades. Paste the function plus its direct callers, not the whole file.
- State the constraint explicitly.
TypeScript strict mode,no new dependencies,must run on Node 18,output must match this exact shape. Unstated constraints are the most common reason a “correct” answer fails in your project. - Include the error verbatim, plus the command you ran. The fix is usually hiding in the stack trace; without it the model guesses.
- Ask for the fix plus a one-sentence explanation of why the bug happened. Skip the explanation and you will not catch the sibling bug yourself next week.
- Run it in your environment before it touches anything real. Run the actual unit test or repro case.
- If the first answer is wrong, do not re-paste everything. Reply with the new error and let the chat compound context. After two or three follow-ups, accumulated context starts to confuse rather than help, so start a fresh chat.
Use Code Interpreter for anything with a file
A trap with regular chat: ChatGPT predicts what your code outputs, and on data it is often confidently wrong. Code Interpreter (available on all paid plans, formerly “Advanced Data Analysis”) fixes this. Upload a CSV or paste a dataset and ask, and ChatGPT spins up a sandboxed Python environment server-side, actually runs the code, and shows you both the real output and the code that produced it.
Use it when the answer depends on data the model cannot eyeball:
- Quick cleaning, charting, or summary stats on a CSV.
- “Does this regex match every row in this file?” — let it run, do not trust a prediction.
- Parsing a log to count occurrences, or reshaping a small dataset.
Two limits to remember (as of June 2026): the sandbox resets between sessions, so download anything you want to keep, and it cannot connect to a live database (Snowflake, BigQuery, Postgres). You upload files; it does not reach into your warehouse.
Note: writing/coding now happens in chat, not Canvas
If you used Canvas for code, note that as of GPT-5.5 it is no longer available in Instant or Thinking. Code now comes back inline as code blocks in the response, with a copy button. There is nothing to migrate; you just copy the block.
Prompt patterns that work
1. "Fix this Python function so it correctly handles
nested dicts. Current code [paste]. Failing test
[paste]. Constraint: no new imports. Explain the
bug in one sentence."
2. "Write a bash one-liner to find all files modified
in the last 24h, larger than 1MB, under ./src.
macOS, not Linux. The find flags differ."
3. "Explain this regex character by character:
[paste]. Where would it fail on unicode input?"
4. "Convert this SQL from MySQL to Postgres syntax.
[paste]. Call out any behavioral differences."Keep these in a personal chatgpt-snippets.md and swap the inputs each time. If you repeatedly do the same kind of work (lots of bash, lots of SQL), put the constraint in Custom Instructions once so you stop re-typing it.
Verify before you trust
The most expensive bugs come from code that reads correctly and runs wrong:
- Did you actually run it, not just read it? For regex and SQL especially, looks-right routinely lies.
- Any imports or APIs you do not recognize? Models past their training cutoff suggest deprecated calls. One search confirms whether the function exists today; for fast-moving libraries, paste the current docs into the prompt.
- For SQL, check the result row count against what you expected before you ship it.
- For regex, test at least three inputs: a clear match, an edge case, and a should-not-match.
Common mistakes
- Pasting huge files. Past ~300 lines the model loses focus. Paste the relevant function plus its callers.
- Omitting the error message. Without it the model guesses; with it the fix is often a one-liner.
- Trusting data answers from regular chat. Use Code Interpreter so the code actually runs.
- Asking it to “review my code” with no goal. You get generic style nits. Ask “find the off-by-one bug” or “where can this throw?” instead.
- Using Instant for real logic. Switch to Thinking for anything past ~20 lines.
- Pasting screenshots of code. Always paste text; OCR-of-code introduces silent character substitutions.
FAQ
- Which GPT-5.5 mode should I use for code?: Instant for one-liners and lookups; Thinking for anything with logic or past ~20 lines. Pro (on the $200 tier) only for genuinely hard problems where a wrong answer is costly.
- What are the message limits?: As of June 2026, Free gets 10 GPT-5.5 messages per 5 hours (then it falls back to a mini model); Go and Plus get 160 every 3 hours; Plus adds up to 3,000 Thinking messages per week.
- Can I paste my whole codebase?: No. Even when it fits, attention spreads thin. Paste the relevant function plus its direct dependencies, or use Cursor / Claude Code, which can read the repo.
- Will ChatGPT train on my code?: Not if you turn off “Improve the model for everyone” in Settings - Data Controls. Defaults vary by plan, and Business/Enterprise/Edu data is excluded by default.
- Why does it suggest deprecated APIs?: Training data has a cutoff. For libraries that move fast, paste the current docs into the prompt so it works from today’s API, not last year’s.