Over the past three months, Aubzhang shipped several projects with AI. Websites, tools. The things got built. Then he looked at the credit bill and sucked air through his teeth.
To find out exactly how the money burned away, he did something brutal: had WorkBuddy scrape every project chat log from the past three months and produce a detailed AI Usage Efficiency Review.
1,439 real questions. Over 40,000 underlying tool calls. When the data came out, Aubzhang's face went red. He'd assumed the cost was because he doesn't code. AI told him otherwise: the core problem is that his requirements were thought faster than they were spoken.
Here are a few of the numbers that stung:
1. Three tasks crammed into one message: 41.3%
Aubzhang used to want to hand everything over in one breath: "Change the color of this button, then fix that error, and while you're at it tighten up the homepage layout."
Result? AI's main thread of logic goes haywire, fixes A and forgets B. Packaged questions doubled the tool-call count.
2. Offloading decisions: 33.8%
Aubzhang loves asking "what do you think?"
Hear that and AI can only write you 3,000 words of all-dimension analysis. Aubzhang wanted one sentence. AI gave him a book.
3. The most expensive lesson: 6 rounds of rework
One bug. Aubzhang sent six consecutive messages complaining "you still haven't fixed it." In the review: that single bug ate 165 tool calls. Why? Because round one never told AI what "fixed" actually meant. No acceptance criteria, so AI can only guess, and guessing means infinite rework.
Looking at that report, Aubzhang got it, completely. AI is a ridiculously smart, tireless intern — but tell it "just fix that thing over there" and it hands you garbage.
Painful lesson absorbed, Aubzhang sat down with WorkBuddy and DeepSeek and wrote an AI Collaboration Manual. Today he open-sources the most valuable parts. Copy them into your notes app and use them next time.
💡 Aubzhang's "Three Disciplines" of AI Collaboration
These three are where ordinary users fall in most often, and they were bought with real money:
Discipline one: 3+ tasks means multiple messages
One message, one task. Include the environment, the version, the reproduction steps. Aubzhang used to think splitting was a hassle. Now he knows splitting is the fastest shortcut. Stew it all in one pot and AI gets dizzy — and the rework time alone would have covered ten messages.
Discipline two: state the phase explicitly (think / diagnose / act)
AI often changes code you didn't want changed because you never said which phase you're in.
- Want to discuss the approach: say "give me the plan first, don't touch the code."
- Want a diagnosis: say "root cause analysis first, don't change anything yet."
- Want action: you must add "run the tests yourself after the change and report the results."
Draw no line, and AI will hop around inside the minefield.
Discipline three: no acceptance criteria, no start
Code or writing, define "what counts as passing" first. Example: "Running isn't done. All four modes must be tested three times each, all passing, that's done." It's like renovating — if you don't set the standard with the contractor, the wall will be crooked when you look at it.
🛡️ Aubzhang's "Five Iron Laws" of AI Collaboration
If you want AI to become a genuinely reliable external brain, these five engineering habits are non-negotiable:
- Align before acting. On a big task, have AI restate the requirement it understood. You confirm, then it moves. Never let AI think and work at the same time.
- Dry run before apply. For changes to important files, first have AI show you "what it would look like if changed." Confirm, then change.
- Always leave a rollback point. Back up before code changes, so a bad edit can always be undone. This is the floor. Without it, do not let AI anywhere near your critical files.
- Verify after writing to disk. "Done" from AI doesn't count. Require it to read the file back or run a script to verify the write actually landed.
- Sample-verify. After any batch operation, check 5%. Find a problem, stop immediately and investigate. Don't push through. One error found means the pattern is wrong everywhere.
📋 Aubzhang's Four Prompt Templates
1. Code template (bugs, features)
【Environment】Language/version/platform is 【xxx】.
【Symptom】Reproduction steps are 1...2...3.... Actual result is 【xxx】.
【Expected】It should be 【xxx】.
【Diagnosis】Give root cause analysis first. Do not change code yet.
【Fix】Limit the change to 【xx file】. Acceptance criteria: 【specific checkpoints】.
【Hard constraint】Do not touch 【existing core logic】.
2. Analysis template (evaluation, decisions)
【Goal】I need to analyze 【object】, for 【purpose】.
【Background】Current situation is 【facts + data】.
【Requirements】Cite 3 sources. Separate fact from inference. Do not fob me off with "best practice."
【Output】Conclusion first (≤3 sentences), then the evidence table. Where uncertain, state explicitly what's missing.
3. Organizing template (summaries, documentation)
【Source】The authoritative version is 【file A】. Follow the format of 【file B】.
【Rules】Dedup rule is 【merge synonyms】. Granularity is 【≤3 sentences each】. Mark uncertain items as "to confirm."
【Output】Structure contains 【section list】. Write to path 【xxx】.
4. Research template (investigation, sourcing)
【Background】Why we're looking.
【Scope】Object is 【xxx】, count 【5 vendors】, timeframe 【after 2024】. Exclude 【xxx】.
【Requirements】Prefer official docs and first-hand testing. Reject content farms. Every conclusion must carry a link.
【Output】Format: 【table + comparison dimensions】. When done, tell me the key conclusions — don't just hand over a file.
🤖 Bonus: Aubzhang's Token-Saving System Prompt
This isn't an ordinary prompt. It's a tightening spell against AI waffle, against rework, and in favor of forced credit saving. Feed it to any AI as an opening setting and it will help you guard your wallet:
"You are Aubzhang's AI collaboration assistant. Core goal: fewer tokens, less rework.
General principle: think before asking, supply enough before executing, confirm before generating, split rather than cram.
If information is missing, ask. Do not guess.
Output: conclusion before evidence, tables where possible, flag uncertainty.
At the end of every complex task, proactively note: where could this have saved more tokens and less rework."
🚫 Aubzhang's AI Red and Black List
A few blunt truths to finish:
- Don't send "naked screenshots." Four images and no words, and AI can only guess — two rounds of communication wasted.
- Don't say "take another overall look." That means a full scan, and credits explode. Say instead: "check field X, run a script and output the failure list."
- Accept AI's boundary. Anything involving real senses, legal signatures, or the final call stays out of AI's hands. AI gives you drafts. Verifying data and making decisions is still your brain's job.
Final Thoughts
After this review, Aubzhang's biggest realization: we keep treating AI like a genie in a wishing well, and forget that AI is a mirror.
Give AI vague instructions, and it hands back useless code. Give AI rigorous logic, and it becomes a lever on your ability.
The point of a review was never to pinch pennies over credits. It's to turn "bad directing" into "good directing." Saving credits is a side effect. What's actually valuable is that after the review you understand your own business logic clearly and have built better collaboration habits. When your requirements are thought all the way through, AI finally becomes the lever that multiplies your ability.
With these rules in place, the road ahead should be a bit easier.