Liz: Session 1 Debrief & Action Plan
Made three configuration improvements during a live show-me-your-AI walkthrough of Liz's actual Claude Pro workflow. Her judgment and verification reflex are already strong; what changed is the organizing layer above the chat box, where the setup had been thinnest. Session 2 will build her first Skill from a repeat-task candidate she flags in the interim.
What the session covered
Liz drove, Steve observed and made three targeted moves. The session anchors:
-
Live walkthrough of Liz's current Claude usage.
The dominant use case Liz demonstrated was Stata
.logoutput reformatted into reader-friendly tables for collaborators. By her own count this used to consume at least half her time. Other use cases shown live: Claude as a sounding board on the basic-science framing for a non-CF bronchiectasis grant; PubMed-style citation searches; reviewer-comment theme extraction. - Default model switched to Opus 4.7 with adaptive thinking in-session. Account had been defaulting to Sonnet 4.6.
-
First Claude Project built live.
"PPI Talk," covering her upcoming presentation. The conversation covered Project Description versus Project Instructions, uploading source files into the project's knowledge base, and the star-the-project sidebar habit.
-
The handoff prompt as a checkpoint pattern was introduced.
Liz, mid-explanation, opened a parallel chat where she had spontaneously run exactly that pattern on her own initiative before the session.
-
Ethics considerations.
Liz identified junior displacement as the most legitimate ethical concern in her own use; environmental impact uncertain but probably not driven by personal use; the "AI lets you be small" framing she got from Berkan's Birmingham AI meetup as her broader read.
Items explicitly parked for session 2:
- Building Liz's first Skill from one of her repeat tasks.
- A short tour of Claude Code as a starting point for the dashboards she has been getting asked to make.
Where Liz stands
The benchmark map. Liz's pushback welcome if anything reads off.
Now
Liz is a daily Claude user with a healthy verification reflex and an unusually intuitive grasp of context-discipline for a first-session client. She had already executed the handoff-prompt pattern on her own initiative before Steve introduced the concept. Her dominant use case (Stata logs to formatted tables) is a clean fit for current Claude capabilities and saves her real time.
What is missing from her current setup is not capability or judgment. Most of her Claude time has been in single-thread chats with no Projects, no Skills, and no model-default control. Each new chat starts cold, with her preferred table format, her notation conventions, her domain context, and her style preferences re-explained from scratch.
Near-term, next 2 to 4 weeks
By late May, Liz's setup looks different than it does today. New chats land on Opus 4.7 with adaptive thinking by default. The PPI Talk Project is the home for that presentation work, with one or two additional Projects spun up for workstreams she returns to repeatedly (the bronchiectasis grant work is a natural candidate). A short running list of repeat-task candidates is accumulating in whatever surface she actually checks, ready to feed the first Skill in session 2.
Further out? See The longer game below.
Top three moves
Ranked by movement-per-effort. Click any move to see the rationale.
1 Hold the line on Opus 4.7 with adaptive thinking as the default
This was done in-session. Opus 4.7 with adaptive thinking is now Liz's default; Steve verified mid-session that new chats land on it automatically, with no Settings change needed. The action this week is just to glance at the model selector when starting a new chat for the first few days, to confirm the default is holding. The cost shows up on harder queries, which take longer because the model is doing more work; the trade is more reliable answers with fewer hallucinations, which matters more for Liz's risk profile than peak responsiveness does.
2 Move ongoing workstreams into Projects
"PPI Talk" is now a Project. The next move is the same setup for one or two other workstreams Liz returns to repeatedly; the bronchiectasis grant work she showed in-session is the obvious candidate. Setup runs about ten minutes per project: create it, write short Project Instructions, drop in the two or three source-of-truth files she keeps re-pasting into chats. Every new chat inside that Project inherits the instructions and files automatically, so she stops spending the first ten minutes re-explaining her own setup. The right time to make a Project is the second time she catches herself re-pasting the same context.
3 Keep a running list of repeat tasks for the first Skill
Skills are saved procedures Claude can invoke when a chat matches their trigger description; they are the next layer above Projects. The action this week is not to build one, but to notice candidates: anything Liz does more than three times that has a predictable pattern qualifies. The Stata-log-to-table reformatting is the leading candidate from session 1; the "extract the common themes from these reviewer comments" pattern is another. Steve will help her build the first one in session 2; the input she needs to bring is a concrete example of the workflow she wants to crystallize.
Steve will send the prompt as a text file. Use it whenever a chat gets long enough that the scroll bar shrinks to a small pill and Claude starts losing thread, or whenever there is an older chat worth pulling into a Project and starting on Opus 4.7. The pattern is the same in both cases: ask Claude to summarize the conversation, then paste that summary as the first message of a new chat. The new chat reads as if the previous one never paused.
Open threads
- Berkan outreach. Steve will reach out in the next week.
- Session 2 scheduling with Liz. Tentative agreement to meet again; no date locked.
-
The dashboards question.
Liz's stated open need (people keep asking her for dashboards and she does not know how to make one) is a natural fit for Claude Code, which sits one or two sessions out from where the setup work needs to be. Parked rather than started.
Commitments
Steve's commitments to Liz
- This brief.
- A handoff-prompt template, sent as a text file.
- Outreach to Berkan in the next week.
Liz's commitments
- Default new chats to Opus 4.7 with adaptive thinking; confirm the setting is sticking for the first week.
- Set up at least one additional Project for an active workstream.
- Keep a running list of repeat tasks as candidates for the first Skill.
- Book session 2.
The longer game
A session 2+ topic, surfaced here for context.
Liz said something in the session that sets the direction this work is pointing:
It's no longer an impediment to be small. AI lets you be small. You can be small while still being big, because you are able to outsource so much of the minutia that you can do a lot more as a smaller company, and you can compete more with bigger people.
Why this is the direction
The mechanical work that used to consume half her time (formatting Stata output into tables for collaborators) has already largely been compressed. Those reclaimed hours open up a wider range of grants she can take on, faster turn-around on consulting requests from colleagues, and the dashboards-and-interactive-artifacts work she has been getting asked for but has not been able to deliver.
The pattern is Liz becoming the AI-fluent biostatistician in her department, with a personal stack of Skills that crystallize her preferred formats, occasional Claude Code work for one-off interactive artifacts, and a working relationship with the department's in-house AI specialist where she requests features rather than only consuming his tools. Each piece is small and incremental. Together they change what she can take on without expanding her hours.
Suggested session 2 agenda
Liz expressed interest in another session and Steve is too. Likely topics, in priority order:
- Build Liz's first Skill from a repeat-task candidate she has flagged in the interim.
- Quick check on the Projects habit. What stuck, what didn't.
-
Brief tour of Claude Code, with the dashboards question as the framing.
If Liz is interested, the first Claude Code project (a small interactive dashboard for one of her current projects) becomes the trigger for session 3. The tour itself stays brief: what Claude Code does, what form the output takes, and the smallest version of a first project that would feel like a win rather than a homework assignment.