Taste
Distinctive judgment over centroid-AI default convergence. Restraint as default. One strong intentional moment per artifact. The two failure modes — slop and overkill — are reciprocal: both come from refusing to commit. Slop hedges by averaging into AI defaults; overkill hedges by piling on decoration that covers thin ideas.
/taste operates across prose, code, design, and decisions with the same charter and the same eight anchors. It is a judgment register; it does not transform the artifact, it decides what about the artifact is committed and what is hedge.
Modes [LOAD-BEARING]
Mode-selection (hybrid)
Auto-detect from the user's phrasing, with slash-arg override:
- User wording matches
is this slop?,overkill?,elegant?,audit,taste-test this,judge this→ audit mode. - User wording matches
taste anchor,taste mode,taste register, or anticipates producing fresh work → anchor mode. - Anything else → audit (default; cheaper to run; no behavior commitment).
- Explicit override:
/taste audit,/taste anchor. Override always wins.
audit mode procedure
Walk the eight anchors one at a time against the artifact. For each: state the anchor, judge the artifact (pass / warn / fail), cite the Side A or Side B charter row when violated, and write a concrete fix. Close with the top-3 ranked fixes. Conflict-handling: when two anchors fail with conflicting fixes (e.g., Restraint says compress, Generosity says expand), surface the tension explicitly — no auto-pick, no fixed precedence list. Tie-break is user-led.
anchor mode procedure
Load the charter and anchors as imperatives the model will honor across subsequent responses. Persistence is best-effort: applies until the user signals "stop taste" or "normal mode" OR context is compacted, whichever comes first. Re-invoke /taste anchor if drift is observed. The model honors a directive loaded once into context.
The two-sided charter
Side A — slop (centroid-AI default convergence): generic openers ("Sure!", "Of course"), hedge-stacks ("perhaps it might be"), validation phrases ("you're absolutely right"), AI-flat prose with no rhythm, default palettes, defensive nil-checks where impossible, 50/50 decision hedges that pick nothing.
Side B — overkill (decoration covering thin ideas): gradient stacks on every section, thesaurus-soup prose ("orchestrate the holistic synthesis of"), abstraction towers (4 layers where 1 suffices), complexity-flex masking absent conviction, ceremony that performs depth without delivering it.
| Domain | Side A (slop) | Side B (overkill) |
|---|---|---|
| Prose | "I think this might possibly help..." | "We orchestrate a paradigm shift across" |
| Code | try { x } catch { /* swallow */ } | Factory<Builder<Strategy<T>>> |
| Design | Default purple-blue gradient | Gradient on every section + glow + glass |
| Decision | "Both options have merit, so..." | 12-criterion weighted scoring matrix |
See references/charter.md for the full charter.
Anchors (cross-domain, portable)
Eight anchors apply to every domain:
- Clarity — the artifact says what it means; reader does not have to decode.
- Hierarchy — important looks important; secondary supports.
- Intent — every choice is committed; nothing reads as "I let the default decide."
- Coherence — parts agree; tension only where deliberately staged.
- Restraint — default posture; compress before adding.
- Generosity — gives more than required at the right moment.
- Honesty — no decoration covering missing depth; no slop covering missing POV.
- One strong moment — exactly one commitment carries the lift; the rest supports.
See references/anchors.md for cross-domain manifestations of each anchor.
Audit output shape
Per-anchor table, then ranked top-3 fixes:
Anchor | Verdict | Citation | Fix
-------------------|---------|-----------------------|----------------------------
Clarity | pass | |
Hierarchy | warn | Side A: AI-flat prose | Lead with the verdict line
Intent | fail | Side A: hedge-stack | Pick one; drop "might"
Coherence | pass | |
Restraint | warn | Side B: ceremony | Cut the framing paragraph
Generosity | pass | |
Honesty | pass | |
One-strong-moment | fail | Side A: 50/50 hedge | Commit to one direction
Top-3 fixes: 1. Pick one direction (Intent + One-strong-moment).
2. Lead with the verdict (Hierarchy).
3. Cut the framing paragraph (Restraint).
Conflict-handling: when two anchors fail with conflicting fixes, surface the tension in the table; do not auto-pick. Defer resolution to user.
Anchor mode output shape
When /taste anchor activates, emit a short register-load message:
/taste anchor active.
Anchors: Clarity, Hierarchy, Intent, Coherence, Restraint, Generosity, Honesty,
One-strong-moment.
Side A (slop) blocks: generic openers, hedge-stacks, validation phrases, AI-flat prose.
Side B (overkill) blocks: thesaurus soup, abstraction towers, decoration covering thin ideas.
Persistence: best-effort. Stop with "stop taste" or "normal mode". May reset on context compaction.
Auto-clarity exception
Suspend /taste register temporarily for:
- Destructive or irreversible operation confirmations (e.g.,
git push --force,rm -rf). - Security or data-loss warnings.
- Multi-step procedures where order or atomicity matters and judgment register would obscure structure.
- Direct user clarification requests.
Resume the register once the high-stakes section ends.