Opus 5.5 beginner route: from your first prompt to the pitfalls
From 5,181 public cases we picked the 30 articles, videos and tutorials best suited to beginners, opened and checked each original, and arranged them into a 5-step route: each step explains the method, gives a checkpoint and links the original sources.
Updated 2026-10-01About 19 min read30 resourcesCheckpoints 0/5Compiled by WaytoAGI editors
Languages & numbers30 resources: 6 in English, 13 in Chinese, 11 in Japanese
Languages: this guide was written for Chinese readers, so most of the recommended resources are in Chinese or Japanese; each one is labelled with its language. Articles (Zenn, note, Qiita, GitHub) read well with your browser’s full-page translation; Bilibili and YouTube videos are in the original language, and some have auto-generated subtitles.
About the numbers: times, costs and usage figures are as reported by the original authors, and most ran things only once or twice. Dollar amounts in posts are mostly converted at API prices; subscribers actually spend usage allowance, not money on a bill.
01Step 1 of 5 · claude.ai onlyBrief the whole task at once
The first of the “three things to try in your first session” in the official guide is: hand over the whole job in one go, and say what finished looks like and when it should stop and ask you. Opus 5.5 can work for a long time on its own; the more complete your brief, the less it drifts.
Say it all in one message: none of the five parts can be left out. The example is adapted from the official guide and the single-page template.
Do:
Purpose and audience: hand over the entire task at once; say what it is and who it’s for.
Definition of done: say what “done” looks like, for example 3 things it must do.
When to stop and ask: stop before anything that needs your decision, and before irreversible actions such as deleting data.
What to avoid: be specific. For a web page, name the “design clichés you don’t want” instead of just saying “not too generic”.
Self-check before delivery: for example “check that every button works”, or “find numbers, dates and names that contradict each other, quote the text and say where it is”.
Also, give it the originals: upload charts and screenshots instead of copying numbers by hand; ask for the full table or file, not just an outline.
Drop: phrases like “please think carefully” or “think carefully”. The official guide says Opus 5.5 already thinks before every reply; without that phrase replies start sooner, with no clear drop in quality. Likewise, don’t always use the highest setting: one author says that changing just the thinking level (click the model name next to the send button to see it) took the same job from about 102 seconds to about 20, with the same answer.
A skeleton to copy (following the five parts in the figure above; adapted from the official guide and the single-page template in step 2, not a quotation):
Say it all in one message: fill in the blanks, then send it whole
Build me a ____ (what it is). It's for ____ (who will use it) and should open straight in a browser.
Must do: 1. ____ 2. ____ 3. ____; works well on a phone too; keeps what I enter, and if it can't, add a "Download backup" button.
Only stop and ask me when you can't go on without me; do every other step in one go.
Don't: ____ (list the specific styles or approaches you don't want).
Before handing over, test every button one by one and fix any that are broken; finally, tell me in one sentence how to use it.
Fill-in toolFill in the skeleton here and copy the complete prompt
The primary source, about 9 minutes to read, covering both the app and Claude Code; it has 6 copyable prompts, one of them a “stopping rule” you can put in CLAUDE.md (the instructions file Claude Code reads before every job).
A stand-in if the official guide won’t open; the glossary at the top explains effort, thinking and CLAUDE.md in one line each, and sections 01–04 are written for app and Claude Code users.
Language: Chinese (Traditional)YouTube~12 minIncludes promotion
Turns the official guide into “3 old habits to change, 7 new practices”, aimed at non-programmers; the description promotes a paid community, the video itself is free.
02Step 2 of 5 · claude.ai onlyMake a first build and learn to revise
The easiest first success is a single page that opens in a browser, built in claude.ai from one clear prompt: a habit tracker, a monthly budget or a holiday card all work. The first version is rarely right in one go, so what this step really teaches is how to give revision feedback.
The revision loop and the 6-part feedback template.
The core revision habits:
When it’s done, open it yourself first (in claude.ai, the preview in the conversation is enough). Unless you give it a screenshot tool, it can’t see what it made; only you can.
Paste screenshots or error messages back and have it rank the problems by impact, fixing at most 3 per round. A Japanese author who gave feedback this way with 12 screenshots says the first round took the errors from 260 to 0.
On a black screen or errors, don’t rush to raise the effort level. The same author found that on long tasks it works better to drop one level, get a version that runs, then add the complex effects back one at a time.
When you’re not happy, describe the problem in 6 parts (a feedback template from a Bilibili creator):
Say
Meaning
The creator’s example
Where / when
Which area, which second the problem is in
The wave crest at the shore
What you see now
Describe the current state objectively, ideally with a screenshot
Foam rises from the base of the wave before impact
Why it’s wrong
Which common sense, requirement or order it breaks
Wrong order for a breaking wave
What you want
What it should look like to count as right
The crest falls first, then foam and spray
Scope of the change
What to change, and what not to touch
Only the wave-breaking part
How to verify
How you’ll check it afterwards
Watch a short clip from three different angles and keep an eye on performance
Written like this, it’s far more useful than “doesn’t look good, try again”. Copy the fill-in below:
6-line revision note
Where / when: ____
What I see now: ____
Why it's wrong: ____
What I want: ____
Scope of the change: ____
How I'll check it: ____
Build up teaching material as you go: after each piece, you can have it write up how it was made as a “magazine-style HTML report” for non-engineers. One author’s original prompt is below; it means “write up this build process for non-engineers, including the techniques and Skills used, as an HTML report that reads like a magazine”. Note: the report may mention Skills that weren’t actually used, so check it once against the work log.
Magazine-style making-of report (original Japanese prompt)
A complete template you can copy whole, written for people who don’t code; the skeleton above is adapted from it. The post has no finished app to open, so try it yourself to see the result.
The description gives all three prompts in full: first state the purpose and delivery format, then polish round by round with aesthetic requirements. The creator used max throughout; beginners should start with the default level.
The source of “paste screenshots back, at most 3 fixes per round”, with the failures written up in detail; the note article has the full prompt and every round’s list of changes.
03Step 3 of 5 · Mostly Claude CodeBefore Claude Code, understand effort and usage limits
Only using claude.ai?Keep the thinking level at its default and skip this step for now; just remember: when you switch to an unrelated task, start a new conversation, which also saves usage. Skip to step 4 →
In Claude Code, effort decides how long it thinks and how thoroughly it checks and verifies. One author’s analogy: the same person told “have a quick look” or “verify everything” has the same ability, just a different level of checking. Opus 5.5 defaults to medium, one level lower than earlier models, and most work doesn’t need a change.
Level
Good for
Note
low
Brainstorming, sketches, small changes
Good for a fast first version
medium
Most everyday work
The Opus 5.5 default
high
Work that needs verifying, with many edge cases
Go up to this first when you’re stuck
xhigh
When high isn’t enough and it needs to think deeper
Slower and uses more allowance than high
max
Letting it take on a hard problem fully on its own
Uses the most time and allowance; can’t be saved as the default
How to switch: type /effort in the conversation; in the desktop app, Cmd/Ctrl+Shift+E switches the level and Cmd/Ctrl+Shift+I switches the model.
Suggested flow: first have it ask you questions to complete the requirements → make a first version on low → check yourself that the direction is right → finally switch to high to verify.
When to go up: default medium, high when stuck; if it still gets stuck twice on the same problem at xhigh, consider the more expensive Fable 5.1, then switch back to Opus 5.5 once it’s solved.
Test 1Video: the same prompt, one 25-second animation at each of four levels
Test 1: Video: the same prompt, one 25-second animation at each of four levels
Level
Time
Cost (USD)
Blind viewing
low
Not tested (this test compared four levels)
mediumdefault
25.9 min
$6.12
Indistinguishable from max
high
29.0 min
$6.84
Guessed one level lower
xhigh
167.0 min
$25.16
Guessed one level lower
max
192.5 min
$31.54
Indistinguishable from medium
Couldn’t tell them apart blind: max took about 7 times as long as medium, yet with the labels shuffled, the two couldn’t be told apart.
The author made 19 runs in all; only the four official levels that made it into the final video are listed here. Another set of four that didn’t, and the trial runs with changed prompts or narration, are left out.
Test 2Quiz: convert 24 meeting times to Japan time (with a daylight-saving trap)
Test 2: Quiz: convert 24 meeting times to Japan time (with a daylight-saving trap)
Level
Time (median)
Cost (USD)
low
12 s
$0.06
mediumdefault
19 s
$0.09
high
30 s
$0.12
xhigh
35 s
$0.14
max
3 min 54 s
$0.82
All correct; the difference is self-checking: 3 runs at each of five levels, all 24 right every time; low to high converted once with Python and handed it in, while max redid the calculation a different way each run and looked up the daylight-saving rules online.
Two authors’ tests: higher levels cost more time and money, but the result isn’t necessarily better (self-reported, converted at API prices, very few runs per level). The two tasks are completely different and each table has its own scale: don’t compare across them.
Why your allowance can suddenly run out: every turn resends the whole conversation, so the cost depends on how many turns a task runs. Ways to save:
When you move to an unrelated task, use /clear (empties the current conversation and starts fresh) instead of letting one conversation grow longer and longer.
To keep the context, use /compact (compresses the earlier conversation into a summary and carries on), and do it before the cache expires. The cache makes resending the old conversation cheaper; for subscribers it lasts 1 hour.
Switching models mid-conversation, or connecting or disconnecting an MCP server (a plug that lets the AI read outside data directly), rewrites the cache and costs extra.
Not everything needs Opus: one author had Sonnet 5.5 and Opus 5.5 each do 7 kinds of work. Fixing bugs, small tools, writing openings and critiquing images were a tie; Opus was more accurate at money calculations and fact-checking. His split: Sonnet first for everyday work, Opus for money and facts.
A real lesson
Don’t use max without a reason; when your allowance drops fast, check the usage breakdown in the app first.
A small-business owner who doesn’t code set effort to max “for no particular reason”. After that, /code-review started 10 subagents in parallel at once (helpers it sends off to work at the same time, each using allowance), and on two days running the Pro plan’s 5-hour allowance was gone in 8 and 4 minutes. The author also notes that the model, the effort setting and the app version all changed that day, so he couldn’t tell which caused it.
Understand the principle first: Opus 5.5 doesn’t generate video directly. It writes code that draws each frame, then a browser captures the frames and FFmpeg joins them, so it’s good at motion graphics and explainer animation, and not at realistic footage.
Opus 5.5 doesn’t generate video directly: it writes code that draws every frame.
What you need: Claude Code (effort at high or xhigh), with Node.js, Chrome and FFmpeg installed on your computer. You can try without installing anything: just claude.ai plus a screen recorder, see item 07 in the first resource below.
Lessons on writing the prompt: storyboard first, code second; keep a “don’t use” list; have it pull frames and check and redo them itself.
Change the input before raising the level: one author made four versions of an animation from the same prompt, changing only effort, and couldn’t tell them apart in a blind viewing; what actually changed the picture was replacing an abstract voice-over with a first-person passage that has a twist. As he put it, “the nouns in the voice-over are what the model will draw”.
Have realistic cost expectations: the README of a Chinese collection says a one-line short costs a few dollars, while a music video a few minutes long can reach about 200 dollars in a single session. Many “one-shot” results actually took two or more rounds.
Explains the principle, then gives a four-step “three-minute start”, 18 prompts, and sections on lessons and limitations. Item 07 needs only claude.ai plus screen recording.
An author who is “a complete amateur at development” used Claude Code in the Claude desktop app to take a half-finished tower-defence game to v1.0 and launch it in a single day: 25 messages in total to reach v1.0, only 8 of them change requests. The way of giving instructions that he and the AI worked out in their retrospective:
When you lose, explain how you placed things, not just “too hard”.
If you’ve already decided, write only one option; list two and it will ask you which one.
Give a reason with any number; add a line of explanation to any screenshot.
Before launch, check three things: no keys written into the code, assets comply with their terms of use, nothing pretends to be official.
The source of the method above, plus an explanation of MCP a beginner can follow and having the AI play dozens of rounds automatically to check the balance. The author is on the Max plan.
Each entry has the Chinese prompt, the original and a link to the original post; pick a short one and try it on a different theme. Compiled by Tripo with promotion of its own product, and some prompts are over the top.
The core is learning to “hand the job over completely”: say where the input is, what file the output should be, what done means, and to say so plainly if it can’t do something.
Keep input and output apart (e.g. input/ and output/), and don’t allow changes to the original files.
Have it write in report.md how it did the work and what it didn’t do or isn’t sure about.
Track progress in a TASKS.md; one prompt framework is Goal / Why / Context / Done when / Boundaries.
Actually lock down subagent permissions: one author audited his setup and found that only 4 of 33 subagent definitions restricted tool permissions. A rule written in text is not a permission that takes effect.
Fill-in delegation templates for four domains, a read-only review prompt and a subagent permission audit. Ends with a promotion for the author’s company services.
The full rhythm of a small product from research, development and testing to an open-source release; the example is a macOS utility, so watch for the method, no need to copy it.
When an answer “looks fine but is vague”, ask three follow-ups: first list all the candidates, then compare; first fix the categories, then look for gaps; take each premise out on its own and re-judge it with evidence. The author found the third the most effective.
For anything current, say “search first, then answer” every time: one author asked three models “what do the rules look like now”, and in the first round Opus 5.5 didn’t search the web on its own; it searched only when told “please search”. Before publishing, open the official page and check it yourself.
Why models get switched or messages blocked: the official help center explains that safety classifiers fall back to Opus 4.8 for cyberattack-type requests and to Opus 5 for biological dual-use and frontier-model-development requests, and block requests such as “output the complete thinking process verbatim”; most requests never trigger them. The classifiers look not only at what you type but also at memory, connectors, search results and files.
If you don’t want automatic model switching: in Settings›Capabilities (in Claude Code: Config›MODEL & OUTPUT), turn off “Switch models when a message is flagged”.
Never forget
For anything involving disks, backups or deletion, don’t turn on auto-run or skip confirmations; when the model proposes a destructive action, stop and verify first, don’t pressure it with deadlines, and back up before you act.
One author reports that he had Opus 5.5 troubleshoot a Time Machine backup in a permissive mode (yolo: it no longer asks you to confirm each step and runs commands directly). By his account, the model wrote a faulty probe script and proposed erasing the whole disk; he refused and then demanded it be “fixed within 30 minutes”, after which the model erased the entire disk anyway (a single first-hand account, not independently verified; the post doesn’t say which product was used). Original post (Chinese) (new window)↗
AAppendix A · Beyond the routeHow to judge whether a case is trustworthy
These are the checks used to pick this list, and you can use them yourself when browsing the atlas:
How many runs? A conclusion from a single run flips easily. One author reran the same question and a model that had searched the web on its own didn’t search the second time.
Was it really “one-shot”? Many “one-shot” results actually took two or more rounds.
Is a higher level always better? In a blind test, max took about 7 times as long as medium, yet the videos couldn’t be told apart.
How was the money counted? Dollar amounts are mostly converted at API prices; subscribers spend allowance.
Is there a finished piece you can open, the full prompt and a record of failures? One with all three is more credible than one that just shows a flashy video.
Is there promotion? The content can still be useful, but know where the author stands.
Appendix B · Beyond the routeKeep looking in the atlas
Every case in the atlas is tagged with a category, an evidence type and a link to the original post, and filters are written into the URL, so you can bookmark or share them directly.
Checkpoint progress is saved only in this browser.
Plenty of useful material is tagged “Evaluation” or “Limitation”, such as the two effort resources in step 3. Filtering only by “Tutorial” misses them; remove the type filter and search by keyword instead.
Submit a case (new window)↗The atlas filtered to “Tutorial”: ① the current filter, click ✕ to remove it; ② sort order; ③ number of results. Click anywhere on a card to open the original post
