Three free chat tools, three different jobs

Most people pick one chat tool and try to make it do everything. Then they get frustrated when it’s great at one thing and mediocre at another. That’s not the tool’s fault — it’s the casting.

Different jobs need different strengths. The tool that’s great at finding sourced research is not the same as the one that holds a long, calm thread about your sleep. And neither of them is the right one for “turn this into a 14-day plan I’ll actually follow.”

three jobs, not three apps

Think in jobs. Research, ledger, protocol — discovery, memory, execution. Splitting them keeps your data clean, your sources honest, and your plans realistic. You can technically use one chat tool for all three, and many people do. The point is the role, not the brand.

research: find, vet, and cite

The research helper has one job: bring back sourced answers you can verify. You want footnotes you can click on and primary sources, not vibes. Right now Perplexity is the cleanest experience for this; Gemini’s search-grounded mode is the strongest free alternative. Either one beats “asking ChatGPT and hoping the citations exist” (sometimes they don’t).

ledger: a long, private memory you control

The ledger holds your timeline. Bloodwork from January, sleep average in February, the supplement you started in March, the niggle that came back in April. Long context matters here, because your week makes more sense in the context of your year. Gemini’s long-context mode is excellent for this. So is a Google Doc you paste into the chat once a week. Both work. The fancy one isn’t always the right one.

protocol: turn insight into a small experiment

The protocol helper takes the first two and writes a small, repeatable plan. A morning routine. A two-week test. A list of three questions to bring to your physio. ChatGPT is steady at this — it follows instructions reliably and gives you the same shape of answer when you ask twice. That consistency is what makes a plan something you can actually follow.

• Research helper — citations you can click; short, verifiable answers; source snippets
• Ledger helper — long context; export buttons that work; doesn’t forget last month
• Protocol helper — reliable instruction-following; same shape of answer twice; doesn’t suddenly get creative

what to do when one of them changes

Models update. The roles don’t. So when your favourite tool ships a weird new behaviour, you don’t need to panic. You triage.

1. Detect — notice which job is broken. Bad citations? That’s research. Forgot what you said last week? That’s ledger. Plans are now creative essays? That’s protocol.
2. Isolate — map the failure to one role. Don’t change the whole stack.
3. Swap — pick another tool that does that role well. The bar is “does this job” not “has the most features.”
4. Test — run three real questions through it. Check the answers, the sources, the consistency.
5. Lock it in — write a one-line note in your ledger about which tool you now use for which role. Future you will thank present you.

That sequence is small and reversible. It turns tool drama into a systems problem instead of a panic.

Most fixes are tiny. If research starts hallucinating sources, swap to another citation-grounded tool. If ledger gets forgetful, paste your data into a long-context model and ask for a timeline. If protocol gets weirdly poetic, prompt it harder for a numbered list. The 3-Layer Stack stays intact.

The failure modes are predictable. Research drifts when it strips context or invents sources. Ledger truncates when context windows are short or the chat is too old. Protocol drifts when the prompt is vague or the model gets “helpful” and adds steps. Knowing the modes makes them spottable.

“Good tooling makes mistakes visible. Then you can fix them without second-guessing the whole method.”

Honest read on the evidence. Strong: citation-grounded search beats free-form chat for accuracy on factual health questions. Promising: long-context tools genuinely help people notice patterns across months of their own data. Anecdotal: a templated “protocol” prompt improves adherence — reported a lot, formally studied a little. Treat each accordingly.

Pick tools by job, not brand. Care about export buttons, source links, and consistent outputs. That’s how you avoid being held hostage by whichever model has the best haircut this quarter.

One method. Three jobs. Swap parts. Keep the method.

You don’t need the single best chat. You need clarity about roles and a stack you can swap parts in and out of without rebuilding your whole life. That’s how AI tools become useful, not brittle.