Liz: Session 2 Debrief & Action Plan
Liz opened session 2 by announcing her promotion to full professor. (Congrats!) To resolve her project-organization friction, we offered a clean working model: a project is a workstation, and each chat is a task inside it. We also tightened her model-selection habit, named her first Skill candidate (the Stata log-to-table job), and set a two-week cadence for session 3.
What the Session Covered
-
Project sprawl, diagnosed
Liz's stated friction was losing track of which project held what, because her studies are interrelated. The fix was a mental model: a project is a workstation that holds the files and context for a kind of work, and each chat is a task inside it. Project memory now reads across the chats in a project automatically, and Custom Instructions can be set per-project or globally.
-
Retitle the project or start a new one?
Liz had built a consortium presentation in one project, then needed to write the manuscript from the same material, and renamed the project rather than starting fresh. That was the right call: when the underlying data and context are the same, keep it in one project and retitle as the work evolves.
-
Cross-project referencing
Liz asked whether one project can look into another (her rat MUC5BC work versus her ferret-safety work, overlapping but distinct). Today the answer is to keep each project focused on its own data even when the output format is identical, and bridge with duplicated files or a reusable prompt block rather than a live cross-project link.
-
Model selection, tightened in-session
Liz had reverted to Sonnet 4.6 on her own because she was burning through credits. We reframed model choice as a per-task decision rather than a single default, and switched the picker to Opus 4.8 with Thinking on for the accuracy-critical work.
-
Workflow excavation
The Stata log-to-table reformatting job surfaced again as her dominant repeat task and the clear first Skill candidate. Liz also walked through her department's custom literature-search tool, which returns a short review with real PubMed citations, and described its companion IRB-protocol drafter.
-
Ethics line, restated
Liz draws a firm boundary: her own documents go into AI freely, but she will not upload a colleague's unpublished work. Anything built in this engagement respects that line.
Where Liz Stands
Steve's read on where things stand. Pushback welcome if anything looks wrong.
Now
Liz is a fluent daily user whose judgment and verification reflex were already strong coming into the session. What had been missing was the organizing layer above the chat box. After session 2 the project model is in place: she understands a project as a workstation, knows to retitle rather than duplicate, and has a working answer for the cross-project question that was nagging her.
Her model selection is no longer a single sticky default she is afraid to touch for credit reasons, but a choice she makes per task.
Near-term, next 2 to 4 weeks
Her active studies live in their own projects, each with a short instruction block and its two or three source-of-truth files, so a new chat inside a project starts warm instead of cold. Accuracy-critical and data-heavy work runs on Opus 4.8 with Thinking on; lighter writing and synthesis drops to Sonnet 4.6 to conserve credits. A running list of repeat tasks is accumulating, with the Stata log-to-table job at the top, ready to become her first Skill in session 3.
Further out? See Where This Is Going below.
Top Three Liz Moves
Ranked by payoff for the effort.
1 Match the model to the task, with Thinking on
Stop treating the model picker as one default to set and forget. Use Opus 4.8 with effort set to Extra for the accuracy-critical and data-heavy work, where a wrong number or a dropped row actually costs something: the Stata analyses, the table generation, the grant text where the statistics have to be exact. Max sits above Extra for the rare make-or-break run, but it burns credits fast; Extra is the everyday setting. Use Sonnet 4.6 on High for ordinary writing and synthesis, which is faster and lighter on credits. Keep Thinking on in both cases.
The credit anxiety that pushed everything onto Sonnet goes away once the expensive model is reserved for the work that warrants it rather than running on every casual query.
Opus 4.8 for accuracy-critical and data-heavy work: set effort to Extra, Thinking on.
Sonnet 4.6 for writing and synthesis: set effort to High, Thinking on.
2 Move ongoing workstreams into Projects, one per study
A project is a workstation. Give each ongoing study its own project, write a short instruction block, and drop in the two or three files Liz keeps re-pasting. Every chat inside that project inherits the instructions and files, and project memory carries context across chats automatically, so she stops re-explaining her own setup at the start of each thread.
Retitle a project as the work evolves rather than starting a new one when the data is the same, the way she already did moving from the consortium presentation to its manuscript. 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 Skill candidates
A Skill is a saved procedure Claude runs when a chat matches its trigger. The action over the next two weeks is not to build one, but to notice candidates. Two kinds qualify: anything Liz does more than a few times with a predictable pattern, and any session where Claude produces an output she really likes and wants to reproduce for other studies. Either way, the thought to capture is "turn this into a Skill." The Stata log-to-table reformatting is the leading candidate, and it is the job slated for session 3.
What she needs to bring is one concrete example of the workflow start to finish, a representative log file and the table she wants out of it, so the Skill can be built against a real case rather than a description.
Open Thread
-
First Skill substrate
Liz keeps the running list of Skill candidates: repeat tasks, plus any output good enough to want again. The strongest example becomes the session-3 build.
Commitments
Steve's commitments to Liz
- This brief, sent with a text so it does not get buried in personal email.
Liz's commitments
- Set the model per task: Opus 4.8 Extra for accuracy-critical and data-heavy work, Sonnet 4.6 High for writing, Thinking on.
- Give each ongoing study its own project with a short instruction block and its source files.
- Keep the running list of Skill candidates, with one concrete Stata log-to-table example ready for session 3.
Where This Is Going
A session 3+ topic, surfaced here for context.
Most of the mechanical work that used to eat half Liz's time is already off her plate. The next layer up is durable: a personal stack of Skills that lock in her preferred formats so she stops re-teaching Claude her conventions, and projects that hold each study's context.
The third piece is different in kind: a working relationship with her department's in-house AI specialist. The clearer Liz is about what she wants AI to do for her, the more she can request features rather than only use the tools he ships.
Taken together, these changes let Liz take on more without adding hours to her day.
Suggested Session 3 Agenda
Liz self-scheduled session 3 on a two-week cadence. Likely topics, in priority order:
-
Build the first Skill, live
Build Liz's first Skill from the Stata log-to-table job, against the concrete example she brings.
-
Turn on the PubMed connector
A two-minute upgrade. Anthropic now publishes an official PubMed connector, so Claude can search live PubMed records directly and cite real PMIDs instead of working from memory.
-
Check the Projects habit
What stuck, what did not, and whether the per-study structure is holding up against her real workload.