Everything your team needs to run the Research Workspace, in one place.
How the three skills work, how to run a cycle, where every file lives, and how to stand up a new client. Bookmark this page, it's your home base.
The system at a glance
The Workspace turns raw census data into the inputs for your CareConscious report. The tool does the heavy lifting and lays everything out neatly, but your team makes every important call. Nothing moves to the report without human eyes.
Three skills, run in order
Themer codes the comments, PII Scrubber makes a safe-to-share copy, Analysis stages the findings. Each one ends in a human checkpoint.
Two streams of data
The comments (open-ends) and the numbers (Forsta crosstabs) travel separately and only meet at the final Analysis stage.
It all runs in Claude Code
Everything runs in Claude Code in the Claude desktop app on the ELIXR Team plan. Claude does the judgment and runs built-in scripts for the heavy, repeatable steps; files are read and written through Google Drive for Desktop.
What each skill does.
Run them in order. Each confirms its inputs at Step 0, stops at its own human checkpoint, and emits a plain-language Results Summary at the end.
Themer
Reads the open-ended comments and tags each one with themes and a sentiment (what Nichole used to do by hand). It proposes the theme list, never finalizes it on its own.
- In: the open-end export; the Care Conscious Theme Seed Library
- Does: Pass 1 drafts the taxonomy + sentiment rubric; Pass 2 codes every comment against the locked list
- Out: Pass 1 draft, coded sheet, theme summary
PII Scrubber
Makes a safe-to-share copy by removing anything that could identify a person, while keeping the substance. The rule of thumb: redact who someone is, keep what they said. Over-redaction is the default.
- In: the Themer's coded sheet; the redaction lists + per-client known-list
- Does: Pass 1 mechanical scrub; Pass 2 catches identifying combinations and logs every removal
- Out: a QA workbook, then the final client asset
Analysis
Brings the comments, the numbers, and the Intake_Library together and stages the findings for the report. It does the bookkeeping so your judgment lands earlier and faster, and it never replaces your call.
- In: coded sheet, Forsta crosstabs, Intake_Library, Analysis_Research_Focus_Guide
- Does: labels each finding (Confirmation / Sharpening / Tension / Money Insight / Silence), scores confidence, stages the Top 10
- Out: staged findings Sheet + topline brief Doc
Always flagged for you: every Tension (a finding that contradicts what you knew) gets a mandatory review: the report cannot ship a contradiction without your stated point of view. Sentiment helps pull supporting quotes; it never ranks a finding.
How to run it, start to finish.
You don't paste instructions each time. In Claude Code, run the three skills in order, finishing each checkpoint before starting the next. The lines below are all you have to say.
.claude/skills folder) is installed in Claude Code. Only the Maintainer needs this setup.Themer
Or, for finer control:
PII Scrubber
Or, for finer control:
Analysis
Or, for finer control:
For mid-cycle steers, update the Analysis_Research_Focus_Guide and re-run Analysis. At the checkpoint the team decides which steers become permanent (promoted into the Intake Library, which Claude drafts and the team approves). There is no separate Cycle Adjustments Log.
The Drive is the system of record.
All client and build content lives in the ELIXR Shared Drive (org-owned, survives departures), not a personal Drive. The skills find their inputs by standard name inside one client folder, so you never edit a skill to onboard a client.
The workspace folder
ELIXR-Research-Workspace/
├── CLAUDE.md ← the one file you point at each client
├── README.md ← how to run it, in plain language
└── .claude/
└── skills/
├── themer/ (SKILL.md + scripts)
├── pii-scrubber/ (SKILL.md + scripts)
├── analysis/ (SKILL.md + scripts)
└── shared/ (shared helper scripts)This whole folder lives in Drive and syncs to the Maintainer's computer through Google Drive for Desktop. The client's data stays in its own engagement folder, separate from the workspace.
The engagement folder
Each client has one engagement folder: 2026 - <Client> Engagement 1/. Put this folder's URL in the engagement-folder line of CLAUDE.md. It has two working subfolders:
- Advisory & Analysis/: the knowledge the team curates, the Intake_Library, Theme_Taxonomy, and Research_Rules.
- Research & Census Analysis/: the cycle steers and raw data, the Analysis_Research_Focus_Guide at the root, plus three subfolders:
Data Tabs/·Claude Outputs/·Archive/
Naming standard
Swap <CLIENT> for the client's short code (e.g. MSC) and <N> for the cycle number. These are the exact document titles.
How Claude resolves a file
A native Google Sheet wins over a same-named .xlsx (keep convert-on-upload ON); the Forsta crosstabs are the exception, staying read-only .xlsx. Only the current version lives in the working folder; superseded ones go to Archive/. For dated files, use the most recent. At Step 0 every skill states what it found and stops if anything is missing or ambiguous.
Onboarding a new engagement.
You don't rebuild anything to onboard a client. You're giving the existing system a new folder to point at. Five steps:
Make the new client's engagement folder in Drive
Drop in the new client's files, each saved into the right subfolder
Rename each file to the standard
Point the workspace at the new client
Confirm the folder is synced
That's it. You can start the run order (Themer, then PII Scrubber, then Analysis). You never touch the skills or scripts to onboard a client.
Updating a skill (maintainer only).
Skills are edited only by the maintainer (Jess), on purpose. Adjusting how a skill behaves inside a single Claude Code session changes only that session; to change it for everyone, the maintainer edits the skill file in the workspace and it syncs to the team automatically.
1 · Edit the skill
Edit the SKILL.md (and its scripts in the skill's scripts folder) inside the workspace's .claude/skills folder, keeping the YAML frontmatter (name, description, when_to_use) intact at the top. Keep the name lowercase-with-hyphens. Each skill now includes pinned Python scripts, so a behavior change may touch the SKILL.md, a script, or both.
2 · Save & sync
Save the updates in the workspace folder. Because the workspace lives in Google Drive for Desktop, the team receives updates automatically.
Safeguards & conventions.
These run automatically, so the deliverable is never built on partial or unsafe data.
Validation gate
A run will not build its deliverable until every response is processed. A failed or missing item is left for an automatic re-run, never silently coded or under-redacted. Code 0 = blank / non-response, kept separate from a real theme and from a processing failure.
Privacy defaults
Over-redaction is the default in the client copy (analysis is unaffected, it runs on the internal text). Business-unit and location demographic fields are excluded from the client file and kept in the internal files.
Quant rules
Percentages tab only; any group of 5 or fewer people is suppressed; a single 95% significance gate; cells that are true by definition are skipped.
Analysis guardrails
The master cut is client-specific and resolved each cycle (it is not hard-coded; Hourly vs Salaried was the MSC example); a structural absence can itself be the finding; wins worth protecting are flagged as such; every Tension carries a mandatory-review flag; sentiment is context for retrieval, never a ranking weight.
Pinned scripts
The mechanical steps (the PII scrub, the schema assembly, and the completeness checks) run as fixed scripts, so the format is the same every run and a run stops rather than ship partial data.
How feedback persists (and what chat does not remember).
Claude does not carry feedback from one chat to the next, or from one client to another. Feedback only persists if it is written to a durable file. There are two layers, with different reach:
- Per-client (stays with that client and carries to its next cycle, on human approval): Themer theme edits land in
<CLIENT>_Theme_Taxonomyand the Themer Feedback Log; PII names and false-positives land in<CLIENT>_Research_Rules; Analysis durable facts land in the Intake Library, and confidence-tier overrides in its override log. These are client IP and never cross to another client. - Across all clients: a theme pattern that recurs across clients is promoted (human-gated) into the Care Conscious Theme Seed Library; a universal redaction rule lives in the shared scrub script that every client uses; Analysis carries nothing cross-client by design. Changing a standing rule or any skill's logic is a deliberate Maintainer edit (logged in the Skill Machinery Ledger), never an automatic result of chat feedback, so one offhand comment never silently changes every client's run.
Rule of thumb: if you want a change to stick, write it down. A change for one client gets approved into that client's knowledge files; a change for all clients goes to the Maintainer.
The handoff documents.
The complete references behind this page. Open these for the line-by-line detail.