The 30-Minute AI Audit: What to Keep, What to Kill, What to Try
A practical workflow for triaging your AI tool stack every 90 days. Includes the prompt that runs the triage, the scoring rubric, and the decision matrix.
The problem the audit solvesThe AI tool landscape changes every quarter. Models that were top of the pack six months ago are mid-tier now. Tools that launched last month get acquired and shut down. Your tool stack from January is not the right tool stack for July, and if you are not actively triaging, you are paying for tools you should have cut and missing tools you should have tried.Most people do not triage because triaging takes time. The triage itself becomes another thing to do, and the thing you were triaging is supposed to be saving you time, so the irony compounds. The fix is to make the triage itself fast and repeatable. That is what this post is.The 30-minute shapeThe audit has four steps. Total time: about 30 minutes if you have a list of tools ready, 60 minutes if you have to assemble the list first. The first time you do the audit, plan for 60. After that, 30 is realistic.List every AI tool you currently pay for or use regularly. (5 minutes) Free tools go on the list too, if you use them. The list is the input to the next step.Run the triage prompt against the list. (10 minutes) The prompt returns a structured comparison and a recommendation per tool. You do not have to take the recommendation. You do have to read it.Score each tool on three dimensions. (10 minutes) Frequency of use, quality of output, switching cost. The score is a number 1-5 on each, summed. Anything below 9 gets a decision: keep, kill, or try.Make the cuts, schedule the tries. (5 minutes) Anything that scored below 9 and does not have a "I will use this within 30 days" date gets canceled. Anything in the "try" bucket goes on a 30-day calendar reminder.That is the audit. The prompt is the part that does the work. Let me show you the prompt.The triage promptDrop this into your LLM of choice. Replace the bracketed list with your actual tool list. Read the output. Do what it says, with the scoring rubric below as a sanity check.
GOAL: Triage a list of AI tools I currently pay for or use regularly.
Return a structured comparison and a per-tool recommendation.
AUDIENCE: Me, a person who uses AI tools for work, has limited attention
for tool sprawl, and would rather pay for fewer tools that
work than many tools that kind-of-work.
FORMAT: One table, then per-tool sections.
Table columns: Tool | Cost/month | Primary use | Verdict
(KEEP / KILL / TRY / REPLACE).
Per-tool section: 2-3 sentences on what it does well, what
it does poorly, and what to do next.
CONSTRAINTS:
- "Verdict" must be exactly one of: KEEP, KILL, TRY, REPLACE
- A tool is KILL if the cost is greater than zero and the
use frequency is less than once per 30 days.
- A tool is KEEP if the use frequency is more than once per
week and the cost is justified by the use.
- A tool is TRY if it has not been used in 30 days but has
a realistic use case in the next 90.
- A tool is REPLACE if a free or cheaper alternative would
do the same job.
TONE: Direct, no hedging, no "it depends" verdicts.
Here is my current tool list:
[PASTE YOUR LIST HERE, e.g.
- Claude Pro, $20/mo, daily writing and code
- ChatGPT Plus, $20/mo, weekly brainstorming
- Cursor, $20/mo, occasional code editing
- Midjourney, $10/mo, used 0 times in 60 days
- Perplexity Pro, $20/mo, used 0 times in 90 days
- ElevenLabs, $5/mo, monthly podcast intro
]
That is the prompt. The constraints do the actual decision-making. The model does not have to guess whether a tool is worth keeping; the rules are baked in. The verdict is one of four values, and the criteria for each verdict are explicit.The scoring rubric (your sanity check)The prompt is fast, but it is not authoritative. The model can miss context you have. The scoring rubric is your 10-minute override. Score each tool on three dimensions, 1-5 each:Frequency (1-5): How often do you actually use it? 5 = daily. 4 = weekly. 3 = monthly. 2 = a few times a year. 1 = less than that.Quality (1-5): When you do use it, how good is the output? 5 = better than any alternative. 4 = at least as good. 3 = fine. 2 = mostly frustrating. 1 = I dread using it.Switching cost (1-5): If I canceled this today, how much would it cost me to come back? 5 = high (e.g. trained on my voice). 1 = none (re-subscribe in 30 seconds).Sum the three scores. Anything below 9 gets a decision. Anything at 9 or above stays. The decision is yours, but the rule of thumb is: if the sum is below 9 and the tool costs money, kill it or replace it. If the sum is below 9 and the tool is free, kill it anyway (unused free tools are attention cost, not money cost).What the audit produces, in practiceHere is the result from the last audit I ran, with the tool names redacted to the categories:Kept (score 9+): The tool I use daily. The tool that has a year of my voice in it. Total: 2 tools.Killed (score below 9, cost > 0): 3 tools I had not used in 60-90 days. Combined cost: $50/month. The audit was right; I have not missed any of them.Try (score below 9, cost = 0, plausible use case in 90 days): 1 tool I had been meaning to try for six months. Scheduled for the next 30 days. The scheduling is the part that converts "I should try this" into "I tried this."Replace (a free or cheaper alternative exists): 1 tool I was paying $20/month for, where the free tier of a different tool would do the same job. Switched in 5 minutes. The free alternative is fine. The $20 was for a feature I never used.Net result: $70/month saved, one new tool on the calendar to try, and a 4-tool stack instead of a 9-tool stack. The stack is smaller, the bills are smaller, and the question of "what should I use for this" is much faster to answer.When to do the auditEvery 90 days. Set a calendar reminder. Do not skip the reminder, the same way you would not skip a smoke detector battery check. The cost of the audit is 30 minutes. The cost of skipping the audit is the slow accumulation of unused tools, which costs money and attention over the year.If you are starting from scratch and you have no AI tools yet, the audit is not for you. Pick one or two tools that solve a real problem you have today, and use them until you have a reason to do the audit. The audit is for people who already have a stack. If you do not have a stack, you do not have a problem yet.The fallbackIf the model gives you verdicts that feel wrong, the score is the override. The prompt is a starting point; the rubric is the source of truth. The model can hallucinate features the tool does not have. The model can miss a tool you actually use. Your score is the ground truth, because you have the data the model does not.If the rubric gives you scores that feel wrong, the audit is not the problem. The problem is that you have not been paying attention to what you use. That is fine. The audit is the first time you are paying attention. The second time will be easier.The honest partI have run this audit four times in 18 months. The first time, I cut four tools and tried one. The second time, I cut one and tried two. The third time, I cut one and tried none. The fourth time, I cut none, because the stack had stabilized.The audits have saved me about $1,800 over 18 months, and they have not cost me anything I needed. The stack I have today is the stack I would have ended up with if I had been paying attention from the start. The audits are how I retroactively paid attention. That is the part I would tell past-me if I could: do the audit every 90 days, and you will end up with the right stack faster than you would have without it.Want the full triage system?The Blog Writing Factory has a "research prompt" that is built for exactly this kind of triage. Feed it a list, get back a structured comparison, decide in 20 minutes. The triage prompt in this post is a smaller version of that prompt.Buy on Gumroad. $14
Send me the rough edges if you try it. I read every message.