affinity — Signal affinity diagram

Cluster the project's Signals/ into a handful of product-specific thematic groups — the affinity-diagram step of research synthesis. The grouping is written into each signal's frontmatter (affinity: '<group>'), so the browser app renders it as an editable board and every claim stays traceable.

Trigger

/affinity            ← group all signals (first run) or re-group the unlocked ones
/affinity --suggest  ← propose changes only; write nothing

The groups are dynamic — derive them from THIS product

Do not reuse a fixed taxonomy. Read Product Context.md and the signals, then invent groups that fit the product:

• a music streaming service → Recommendations & algorithm, Library & finding music, Playback experience, Pricing & value, Competition & switching, Accounts & sharing, Input & accessibility, Social & identity …
• a car maker → by vehicle segment, customer type (fleet / sole-trader / private), ownership stage …
• a B2B SaaS → by job-to-be-done, team role, lifecycle stage …

Aim for 5–9 groups: enough to separate real themes, few enough to stay legible. A theme with one lonely signal is usually a sub-point of a bigger one — fold it in unless it's genuinely distinct.

The hard rule — never overwrite human curation

Each signal may carry affinity_lock: true. That means a human placed it (via the app's Edit mode or by hand). You must not change affinity: on a locked signal. Instead:

• Leave it exactly where it is.
• If the data suggests it belongs elsewhere, say so in your report as a suggestion ("Signal X reads more like Group Y — want me to move it?"), and act only if the user agrees.
• Locked signals still inform what the groups ARE — read them, just don't move them.

Unlocked signals (no affinity_lock, or false) are yours to (re)assign freely.

Workflow

1. Read Product Context.md + every Signals/*.md (use the transcript scan headers / signal quotes; don't re-read transcripts).
2. Decide the group set (respecting the themes locked signals already anchor).
3. For each unlocked signal, set affinity: '<group>' in its frontmatter (one group per signal). Leave affinity_lock absent — the app adds it only when a human curates.
4. Report: the group list with counts, which signals moved, and any suggestions for locked signals you'd have placed differently (never applied).
5. First run (no signal has affinity: yet) is fully automatic — assign them all. The app also shows a provisional by-tag grouping until this runs.

Rules

• One signal, one group — affinity is a single value, not a list.
• Group names are short human phrases, product-specific, not tag slugs.
• This never creates or deletes Signals, and never touches anything but the affinity: line of unlocked ones.
• Contradiction-safe: two signals that disagree can share a group (the group is a theme, not a conclusion) — grouping is not synthesis; Correlations remain where you reconcile them.