50 lines
6.7 KiB
Markdown
50 lines
6.7 KiB
Markdown
## Executive summary
|
||
This is a scoping interview for an upcoming webinar on practical AI use, in which the host (SPEAKER_00) probes a power user (SPEAKER_01, "Larysa") for the concrete problems she hits with Claude. The dominant pain points are not hallucination but structural: agents have no useful persistent **memory**, and Claude confidently starts tasks against **integrations that turn out to be unavailable** for her account type — both burning time and tokens. The host's thesis is that the fix is to stop tool-hopping and instead build one consolidated, project-scoped agent "operating system," and to constrain prompts with **skills** rather than expecting a whole project from one or two requests.
|
||
|
||
## Who's who
|
||
- **SPEAKER_00** → The webinar host and AI power user. Advanced Claude Code user who has built a consolidated multi-agent workspace (his "Inspectron" projects) and an interview-analysis system that diarizes speakers and auto-writes conclusions — the same tool that produced this transcript. Stance: strongly pro-Claude, pragmatic, skeptical of one-shot "build everything" demos.
|
||
- **SPEAKER_01 ("Larysa")** → A technical BA/PM and former mobile developer. Hands-on Claude user: opens pull requests, wires Claude to ClickUp (a Jira analog) and Figma, builds clickable prototypes, writes user stories and tests. Stance: enthusiastic but worn down by memory loss and dead-end integrations.
|
||
|
||
## Key themes
|
||
|
||
**Agent memory is the central frustration.** Larysa's core problem is that Claude doesn't carry context between sessions — she explains something one day and it's gone the next, so she re-pays in time and tokens. The host agrees bluntly that memory, as implemented, is a net negative.
|
||
> "Memory is the worst thing agents have — it gives no benefit and confuses users to hell. Why even go there? … The memory exists, but the way it's implemented, it'd be better if it didn't."
|
||
|
||
**Integrations overpromise, then dead-end.** Claude accepts a task and starts working, only to discover mid-way that a connector isn't implemented or isn't available for her (private vs. corporate) account — after she's already sunk half an hour. This is her example of a Teams bot that would watch a work chat and auto-create ClickUp tasks.
|
||
> "What it recommended is available not for private accounts, only for corporate ones — and to get there with it I'd already burned half an hour. That's not hallucination exactly… it just first goes looking for what you don't have. I don't know how to fight this."
|
||
|
||
**Build one system; stop tool-hopping.** The host's main pitch: assemble a personal "operating system" where all your project agents live together and can talk to each other, instead of spinning up Claude in separate folders. He claims it freed up mental overhead he didn't know he was spending.
|
||
> "You build yourself an operating system where you have everything in one place… When I built this thing, my life just split into before and after."
|
||
|
||
**Leave less room for imagination.** Under-specified prompts let Claude drift, and the drift causes large collateral damage; standardized procedures (skills) are the constraint. This is his explicit critique of demos that ask Claude to build a whole project in one or two requests.
|
||
> "The more room for imagination, the more it will exploit it… it finds a way to understand something differently than you imagined, and it just gets carried away — you won't even notice."
|
||
|
||
**Model choice and how far to trust coding AI.** The host runs Claude 4.7 in production, finding 4.8 "too proactive," and argues coding with AI is no longer elite or unsafe — with two exceptions.
|
||
> "I use Claude 4.7 — it does exactly what I need, without the flights of fancy 4.8 has… Code isn't something elite anymore. From 4.6 on, the code is safe enough — though authorization and payments I still wouldn't trust to Claude."
|
||
|
||
**Automated interview analysis.** Both independently landed on recording interviews and processing them, but the host's system adds speaker diarization plus role inference from context, which he says a plain GPT transcription flow cannot do — it only guesses who spoke.
|
||
> "Instead of sitting and taking notes, you just screen-record, then run it through the AI… an hour-long interview took about 10–11 minutes. It splits by voice and pulls the role — Speaker 0, HR Equity Lead; Speaker 2, Senior Programmer — from the conversation itself."
|
||
|
||
## Conclusions
|
||
- The real friction with agentic AI here is structural, not accuracy: (1) no durable memory, (2) integrations that appear available but aren't for the user's account/tooling, and (3) over-eager drift on loosely specified tasks.
|
||
- **Skills** are the host's recommended remedy for the memory and repeatable-procedure problems, and are the one topic he commits to covering for Larysa in the webinar.
|
||
- Teams-based automation is effectively a dead end (Microsoft exposes no good external API); Slack has first-class Claude support from Anthropic, but that requires a paid seat she isn't given.
|
||
- A single consolidated workspace of project-scoped agents — with completion notifications — beats juggling separate tools, both for output and for reduced mental load.
|
||
- AI-assisted coding is treated as production-ready as of Claude 4.6/4.7, with authorization and payment logic as the deliberate carve-outs.
|
||
- Both participants converged on automating interview/meeting analysis; diarization plus role inference is the differentiator over transcribe-then-summarize.
|
||
|
||
## Takeaways
|
||
- Use skills for anything repeatable to work around agent memory limits, rather than re-explaining context each session.
|
||
- Verify integration and account requirements up front (connector exists? private vs. corporate? paid Figma/Slack?) before letting Claude spend time and tokens on it.
|
||
- Prefer browser/web testing over emulator testing with Claude — the emulator got badly stuck.
|
||
- Constrain prompts and don't expect a finished project from one or two requests; leaving room for "imagination" invites drift.
|
||
- Keep authorization and payment code out of Claude's hands.
|
||
- Consolidate into one agent workspace with completion signals instead of tool-hopping across folders and apps.
|
||
- Default to Claude 4.7 for controllable work; only reach for a more imaginative model when the task genuinely needs it.
|
||
|
||
## Open questions
|
||
- How to reliably handle integrations Claude recommends that aren't actually available for the user's account type — left explicitly unresolved ("I don't know how to fight this").
|
||
- Whether a Teams-triggered auto-task bot is feasible at all, given Microsoft's lack of usable APIs.
|
||
- How to get full (editable) Figma project access without a personal paid account.
|
||
- How to structure and schedule the webinar itself — deferred to a follow-up with Yulia.
|