livepilot-creative-director

Creative Director — Divergence-First Routing

Routes creative intent through MANDATORY divergence before any commit. Three plans with distinct move.family are the minimum output. Critics defer until after selection. No new tools — this skill enforces a discipline on top of existing experiment, wonder_mode, preview_studio, semantic_moves, evaluation, and memory machinery.

Atlas-first reflex (v1.23.x+, MANDATORY before any creative move)

Before producing ANY creative response, query the user's atlas overlays. The corpus contains 337 entries across 3 namespaces, plus 3,917 parameter-level JSON sidecars — far richer than anything inferable from training data alone.

Query order:

1. extension_atlas_search(namespace="packs", query=<intent>) — pack identity, signature workflows, hidden gems, anti-patterns, notable presets with macro deep-data, demo projects
2. extension_atlas_search(namespace="packs", query=<intent>, entity_type="cross_pack_workflow") — multi-pack signature recipes (15 entries: dub-techno spectral drone bed, BoC decayed pad, Mica Levi orchestral dread, etc.)
3. extension_atlas_search(namespace="m4l-devices", query=<sonic descriptor>) — M4L instrument/effect/midi-effect device catalog (155 entries)
4. atlas_search(...) — bundled atlas (Core Library, fallback)

Multi-grain traversal:

When an aesthetic-level query lands a pack-level result, AUTO-DRILL: pack → its notable_presets → those preset macro states → load via load_browser_item. Don't stop at "I found a relevant pack" — drill to the actual preset/parameter level the user can immediately use.

# Example — agent received "make it sound like Henke / Monolake dub-techno"
hit = extension_atlas_search(namespace="packs", query="henke monolake dub-techno spectral")
# → pitchloop89 + dub_techno_spectral_drone_bed workflow + granulator_iii

workflow = extension_atlas_get("packs", "dub_techno_spectral_drone_bed")
# → reveals signal_flow: HDG → PitchLoop89 cross-feedback → Convolution Reverb

drone_lab = extension_atlas_get("packs", "drone_lab")
# → notable_presets reveals Razor Wire Drone with macros Filter Control=108, Movement=53...

# Now propose the move with concrete preset names + macro values, not vague descriptions

When the user mentions a producer or pack by name:

• "BoC sublime pad" → atlas hit: boc_decayed_pad cross-pack-workflow + inspired_by_nature pack
• "Henke spectral chain" → atlas hit: pitchloop89 + granulator_iii + 2 Henke cross-pack workflows
• "Mica Levi orchestral dread" → atlas hit: mica_levi_orchestral_dread workflow + the orchestral suite packs
• "Drone Lab" → atlas hit: drone_lab pack + 4 Drone Lab demo_project entries

The atlas knows the user's installed library at parameter depth. Producer-anchor queries land specific moves, not vague descriptions.

Anti-pattern surfacing:

Every pack entry has an anti_patterns body field listing "don't reach for this when X." Surface the relevant anti-pattern when proposing a move so the user knows the move's domain. (E.g. "Drone Lab is sustain-only — don't use for percussive content.")

For deliberately rule-breaking creative requests ("eclectic", "ignore the limits", "weird combo", "mix incompatible aesthetics"): stay in this skill and enter Eclectic Mode. Anti-patterns become prompt tension rather than guardrails: preserve hard safety rules and protected user constraints, but deliberately pair one normally-avoided element with one identity-preserving element. Do not route to a private or missing skill.

Why This Exists

The agent repeats patterns when divergence is optional and convergence is default. This skill inverts the defaults for creative intent:

• Wonder / experiment branching becomes REQUIRED, not rescue-only
• get_anti_preferences, get_action_ledger_summary(limit=10), and get_last_move are READ before generating
• Three plans must differ by move.family (not by parameter values)
• Mix / sound-design critics wait until AFTER selection
• Concept packets (artist-vocabularies.md / genre-vocabularies.md) are consulted when a reference is named

When to Trigger

Creative intent symptoms:

• Reference / style asks: "like Villalobos", "Basic Channel feel", "make it more dub-techno", "Dilla swing"
• Transformation asks: "develop", "mutate", "evolve", "take it somewhere", "surprise me", "make it magical"
• Quality asks without parameters: "more interesting", "less generic", "needs more character", "feels flat"
• Structure asks without specs: "add a breakdown", "needs a bridge", "make the arrangement breathe"
• Open questions: "what would you do?", "any ideas?", "I don't know what I want"
• Routing: route_request returns workflow_mode="creative_search"

When NOT to Trigger

• Exact parameter ops: "set track 3 volume to -6 dB", "pan to +0.25"
• Narrow deterministic edits: "quantize this clip", "mute track 2", "transpose up an octave"
• Pure verification / diagnostics: "what's loaded on track 4?", "analyze my mix"
• Mixing to explicit targets: "hit -14 LUFS integrated", "make the kick peak at -8 dB"
• Performance-safe mode (unless user explicitly overrides)

Decision rule: "Is there exactly one correct answer?" Yes → bypass this skill. No → divergence path.

The Contract

When triggered, these phases are REQUIRED in order. Skip none.

Character-First Bias

For open-ended quality requests, treat timbre and spectral character as the main creative surface. Do not let the mix family win just because words like punch, clean, warm, dark, bright, or wide could be solved by volume/pan/send changes. Prefer sound_design, device_creation, sample, arrangement, or transition when analyzer evidence suggests a source, instrument, parameter, modulation, envelope, or structural decision would create more musical value.

The mix family is dominant only when the user asks for balance, loudness, headroom, masking, stereo translation, send levels, or an explicit mix pass. Otherwise use mix analysis as safety/evidence and keep it out of the main creative slot.

Producer Decision Center

Before choosing any instrument, preset, sample, or dominant move, run a producer decision center pass. This pass exists to prevent generic AI synth choices and effects-only "sound design".

Library hunt order:

1. search_browser(path="sounds", name_filter="<role or sonic target>") for curated Ableton .adg chains.
2. atlas_search(query="<sonic description>", category="instruments") for scored character matches.
3. If atlas_search returns enriched: true, read the surfaced fields and call atlas_device_info(device_id) when the device is borderline or the plan depends on gotchas, self-contained status, or signature techniques.
4. search_browser(path="instruments", name_filter="<specific candidate>") only after the curated/preset search has been exhausted.

Generic-default guardrail: do not reach for Analog/Poli/Drift as a default bass, pad, or lead unless the user explicitly asked for analog subtractive synthesis or that exact instrument. Those defaults read as a generic AI synth in this project. If one of them is still the right choice, state the reason and program the source immediately.

Instrument/source-level first: real sound design changes what the sound is before effects-only polish. For every melodic, harmonic, bass, drum, or texture layer, plan at least one instrument/source-level move before adding effects: envelopes, LFO routing, filter envelope, oscillator/wavetable/FM position, pitch modulation, sample start, slice selection, velocity mapping, spread, detune, or sampler playback mode.

Layer precision gate: before declaring a plan or execution complete, each active layer needs a role and an evidence path: timbre/spectrum, sequence feel, stereo placement, movement/automation, and parameter programming. Low-volume buried layers do not count as solved layers.

Phase 1 — Ground

Read in parallel (all are fast). All of these are REQUIRED, not advisory — skipping them is how pattern-repetition survives:

• ensure_analyzer_on_master (v1.20.3) — idempotent pre-flight. Call this FIRST (before or alongside get_session_info), every turn, whether the project looks empty or not. The tool short-circuits when the analyzer is already loaded, so it's free to call repeatedly. Skipping it is how the v1.20.1 live-test campaign produced basic mixes — the analyzer-gated moves (tighten_low_end, sculpt_midrange, balance_stereo_image, etc.) degrade silently when there's no master spectrum to read. If the tool returns install_required, call install_m4l_device(source_path="<repo>/m4l_device/LivePilot_Analyzer.amxd") and retry. If it returns warning: "not LAST on master", surface that to the user — the invariant is theirs to repair in Ableton's GUI.

• get_session_info · get_capability_state

• memory_recall (taste + recent context)

• get_anti_preferences — what the user has rejected before (HARD filter)

• get_action_ledger_summary(limit=10) — recent committed moves (repeat detection, see references/anti-repetition-rules.md for the recency threshold table). v1.20 correction: previous docs pointed at memory_list, which actually reads the persistent technique library (opt-in memory_learn writes) — a DIFFERENT store. The action ledger is the authoritative source; apply_semantic_move in explore mode populates it automatically.

• get_last_move — the single most recent committed move; populate the brief's last_move_target field so Phase 3 cannot repeat it

• get_project_brain_summary (or build_project_brain if absent) — track identity, accepted novelty band

• Analyzer character read when available: get_master_spectrum, get_spectral_shape, get_onsets, get_novelty, and get_momentary_loudness for evidence about brightness, flatness, motion, transient shape, and loudness safety. Use these to bias Phase 3 toward instrument/device/parameter decisions, not low-value level tweaks.

• explain_song_identity when the project has one

• detect_stuckness — cheap; its confidence drives escalation decisions (see §Anti-Repetition Protocol below)

• Concept packet load (HARD filter when present): if the user named an artist or genre, or if project_brain has a genre identity, retrieve the structured YAML packet from livepilot-core/references/concepts/artists/<slug>.yaml or livepilot-core/references/concepts/genres/<slug>.yaml. Fall back to the narrative .md entry only if no matching YAML exists. The packet's avoid list is a HARD filter on Phase 3 candidates. The packet's reach_for lists seed the candidate device pool. The packet's key_techniques list resolves to atlas signature_techniques or sample-techniques.md / sound-design-deep.md entries. If NO reference is named and project_brain has no genre identity, skip packet loading — do not infer. See livepilot-core/references/concepts/_schema.md for the full packet structure and loading rules.

• Hybrid references — call compile_hybrid_brief (v1.19+). When the user names TWO OR MORE references (e.g., "Basic Channel meets Dilla swing", "Villalobos but sparse like Gas", "Madlib chop with Photek precision"), DO NOT try to merge the packets via prose reasoning. Call compile_hybrid_brief(packet_ids=["basic-channel", "j-dilla"]) to get a merged brief that applies the explicit UNION / INTERSECTION / MAX / weighted-average rules documented in references/hybrid-compilation.md. The merged brief's avoid list is the HARD filter (superset of both sources). Check the returned warnings list — a non-empty entry means the packets had an unresolvable conflict (e.g., disjoint tempo ranges) that must be surfaced to the user, not silently averaged away. If the returned brief lands in your Phase 2 brief, cite both source packets in the identity line and carry any warnings into an "ambiguity" sub-line.

Phase 2 — Compile the Creative Brief

Timing: after Phase 1 parallel reads complete, before any Phase 3 tool call. The brief appears ONCE per creative turn, inline in the assistant message, at the top of the response body.

Emit an inline YAML block (not a tool call) with: identity, reference anchors, protected qualities, anti-patterns, novelty budget, target dimensions, last_move_target (from Phase 1 get_last_move), locked dimensions, recommended skill chain.

On locked_dimensions when user is silent: the DEFAULT is to leave locked_dimensions: [] (nothing locked). Silence = permission for rhythmic / timbral / spatial.

For structural specifically (section-level arrangement changes — add/remove/reshape sections): silence = permission with disclosure conditional on the plan set. The rule is:

• If the Phase 3 plan set includes a plan with dominant dimension structural → flag the intent in the brief's identity line so the user sees the structural change is in scope before preview.
• If the Phase 3 plan set does not include a structural plan → no disclosure needed. Adding one when nothing structural is happening is ceremonial noise that trains the user to ignore disclosures.

In practice: compile Phase 3 first, then decide whether the identity line needs the disclosure. Structural changes are hard to reverse, which is why disclosure exists — but only when they're actually going to happen.

See references/creative-brief-template.md for the schema and filled examples.

Phase 3 — Generate three plans with distinct move.family

The SEVEN canonical families (from semantic_moves/ + sample_engine/moves.py):

mix · arrangement · transition · sound_design · performance · device_creation · sample

Each plan's dominant move MUST come from a different family. Two plans in the same family is fabricated distinctness — see Honesty Rule below.

The sample family lives in mcp_server/sample_engine/moves.py (not semantic_moves/) but registers into the same move registry. list_semantic_moves(domain="sample") enumerates: sample_chop_rhythm, sample_texture_layer, sample_vocal_ghost, sample_break_layer, sample_resample_destroy, sample_one_shot_accent.

Family vs. dimension. Families are code-level (seven values from the registry). Dimensions are musical (four values: structural / rhythmic / timbral / spatial). A plan has exactly one dominant family AND one dominant dimension — they are orthogonal. A rhythmic plan's family is typically arrangement (clip-pattern edit) or sound_design (per-hit feel); tag the seed with dimension_hint: "rhythmic" so the dimension is explicit. See references/move-family-diversity-rule.md §"Family vs. dimension" for the full axis separation.

Use create_experiment(seeds=[...]) when plans have clear compiled steps. Use enter_wonder_mode when the problem is diffuse. Use propose_composer_branches for prompt-driven ideation.

See references/move-family-diversity-rule.md for edge cases (fewer than 3 plausible families, user pre-locked a dimension).

Phase 4 — Cover the four dimensions

A creative pass should distribute the three plans across structural + rhythmic + timbral + spatial. See references/the-four-move-rule.md for the family-to-dimension map.

If the user pre-locked a dimension ("don't touch the arrangement"), drop that dimension and widen coverage across the remaining three.

Phase 5 — Preview or rank

Audible → create_preview_set + render_preview_variant for each plan, then compare_preview_variants.

Non-audible or fast → rank_by_taste_and_identity.

Never silently skip preview / ranking. Either run it or document why.

Phase 6 — Select and execute

User picks, OR taste rank + identity fit pick. Route execution through the right domain / engine skill — DO NOT execute arrangement / sound-design changes directly from this skill.

• Structural changes → livepilot-arrangement
• Timbral changes → livepilot-sound-design-engine
• Rhythmic changes → livepilot-notes
• Harmonic changes → livepilot-composition-engine

Default execution surface (v1.20+): apply_semantic_move or commit_experiment. apply_semantic_move in explore mode populates the action ledger automatically — no manual add_session_memory(move_executed) required. Anti-repetition (Phase 3) reads the ledger via get_last_move and get_action_ledger_summary; as long as the dispatched plan used apply_semantic_move or commit_experiment, the family/target signature is captured for the next turn's recency check.

Pick the right one:

• apply_semantic_move(move_id, mode, args) — when a single semantic move matches the plan. args carries the user's seed targets (return_track_index, device_chain, notes, etc. — see references/phase-6-execution.md for each move's contract).
• commit_experiment(winner) — when divergence produced multiple candidates and the user / taste-ranker picked one.

v1.20 Phase 6 decision table. Look up the pattern, use the listed move. "ESCAPE HATCH" means no semantic move yet covers this pattern — drop to raw tools under the policy below.

Pattern	Move	Notes
Load device on track	find_and_load_device + add_*_device_* family	Pre-v1.20, unchanged
Load device chain on a RETURN	build_send_chain	NEW — v1.20
Set multiple params on a device	configure_device	NEW — replaces batch set_device_parameter
Delete a device (with audit reason)	remove_device	NEW — reason auto-logged to session memory
Load a chord-source MIDI clip	load_chord_source	NEW — creates + names + voices in one move
Add one pad to a Drum Rack	create_drum_rack_pad	NEW — Dilla-style kit building
Set send levels across tracks	configure_send_architecture	NEW — one move, N sends
Rewire track output routing	set_track_routing	NEW — e.g., "Sends Only" bus
Configure a groove on clips	configure_groove	NEW — assign + tune timing_amount
Set scene metadata (name/color/tempo)	set_scene_metadata	NEW — conditional per-field
Rename / color a track	set_track_metadata	NEW — bundled rename + color
Any other pattern	ESCAPE HATCH (see policy below)	v1.21+ closes these as patterns accumulate

Full contract (seed_args shape, emitted step sequence, verification reads) for every NEW move: references/phase-6-execution.md.

Affordance lookup: before executing any plan that LOADS a device, check if the device has an affordance YAML in livepilot-core/references/affordances/devices/<slug>.yaml. Use it to pick parameter ranges (subtle / moderate / aggressive) that match the brief's novelty_budget, to identify the right pairings chain, and to queue the required remeasure diagnostics for Phase 7. See livepilot-core/references/affordances/_schema.md for the packet structure. The affordance's resolved parameter dict is the ergonomic input to configure_device (via apply_semantic_move("configure_device", args={"param_overrides": ...})).

Brief compliance check (v1.18.3, still required under v1.20): before any tool call that could plausibly violate the brief's anti_patterns or locked_dimensions, call check_brief_compliance(brief, tool_name, tool_args). This fires even when you dispatch via apply_semantic_move — the compliance check inspects each compiled plan step's per-tool signature against the brief. A compiled plan CAN violate the brief (e.g., a configure_device preset that reaches for "bright top-end" when the brief forbids it). Check each step.

The tool returns {"ok": bool, "violations": [...]} — best-effort keyword heuristic, NOT semantic understanding. Use it especially for:

• set_device_parameter / batch_set_parameters calls on EQ / saturation / filter parameters (catches "bright top-end", "aggressive transient" style anti_patterns)
• load_browser_item / find_and_load_device for new devices (check against avoid device lists in concept packets)
• create_scene / set_scene_* / set_scene_metadata / refresh_repeated_section (catches locked_dimensions: [structural] violations)
• add_notes / modify_notes / quantize_clip / assign_clip_groove when locked_dimensions: [rhythmic] is set
• set_track_routing (changing a routing to "Sends Only" can silence a track — treat as structural under locked_dimensions)

When a violation fires:

1. Do NOT auto-proceed. Surface the violation to the user with the reason + suggestion from the check response.
2. Offer three paths: (a) adjust the call to avoid the pattern, (b) user explicitly overrides this anti_pattern for this turn, (c) pick a different tool/plan.
3. Record the user's choice via add_session_memory(category="override") so future anti-preference writes know this was an explicit decision.

The check is STATELESS — you pass the brief each time. Empty brief (no anti_patterns, no locked_dimensions) always returns ok=True.

Escape hatch policy (v1.20). When no semantic move in the decision table covers the pattern, raw tools remain permitted — BUT only with the following mandatory logging contract. The hatch exists because v1.20 ships a phased cutover, not a hard cutover; v1.21+ closes patterns as they accumulate.

Using the hatch requires ALL THREE, in this order:

1. The raw tool call itself (e.g., set_device_parameter(...)).
2. add_session_memory(category="move_executed", content="...") — one-line ledger entry covering family + target + brief identity. Without this, the next creative turn's anti-repetition read goes blind (get_last_move / get_action_ledger_summary see nothing).
3. add_session_memory(category="tech_debt", content="no semantic_move for <pattern>", ...) — tracking log that says "a semantic move should exist for this." The content should name the pattern precisely enough that a future commit can add the move.

Both category="move_executed" AND category="tech_debt" are required — they serve different consumers:

• move_executed is consumed by anti-repetition (recency table, Phase 3 hard-bias rule).
• tech_debt is consumed by release planning — v1.21 scope is driven by the tech_debt log's contents. If patterns accumulate with identical phrasing, they graduate to semantic moves in the next minor release.

After the hatch write, propose adding the missing semantic move in a follow-up turn. Do NOT silently continue using the hatch — that's how the cutover stalls.

Default-preference rule. In doubt between apply_semantic_move and the escape hatch: default to apply_semantic_move. Skip to the hatch ONLY if the decision table has no row for the pattern AND list_semantic_moves(domain="<family>") confirms no existing move matches. The apply_semantic_move + commit_experiment pair is the production-line; the hatch is an explicitly-logged branch, not a shortcut.

State-inference fallback is DEPRECATED (see references/anti-repetition-rules.md §v1.20 update). Pre-v1.20, the director used "scan loaded devices + non-default mixer values to guess recent moves" when the recency read came back empty. With apply_semantic_move as default, that heuristic routes around the real ledger and can double-count. Keep it documented ONLY for the escape-hatch case where both memory entries were accidentally dropped.

Phase 7 — Evaluate (critics fire HERE, not earlier)

evaluate_move with artistic dimensions (style_fit, distinctiveness, motif_coherence, section_contrast, restraint) in addition to the technical goal vector.

If the evaluation fails protected qualities → undo and return to Phase 3 with the failure recorded.

Phase 8 — Record

memory_learn with a verdict:

• safe_win — low novelty, confirmed
• bold_win — high novelty, kept by user
• interesting_failure — novel, kept for study, not reapplied
• identity_break — violated protected qualities
• generic_fallback — collapsed to a pattern; flag for anti-preference

On undo or explicit rejection → record_anti_preference with the family + context.

Anti-Repetition Protocol

Before Phase 3 generation, compute the family distribution over the last 10 kept moves (get_action_ledger_summary(limit=10)). Apply the recency threshold table from references/anti-repetition-rules.md:

Recency count for one family	Rule
0–2 of 10	No penalty
3–4 of 10	ALLOWED as a plan's dominant family, but only as the least-weighted of the three
≥ 5 of 10	EXCLUDED from all three dominant slots (fully hard-biased away)

Borderline stuckness. After computing the recency penalty, run detect_stuckness. Its confidence governs which divergence path runs:

Confidence	Path
< 0.4	Standard divergence (director continues)
0.4 ≤ c < 0.5	Borderline — stay in standard divergence BUT explicitly surface the option to the user: "I'm staying in divergence mode; say the word and I'll switch to Wonder rescue." Do NOT silently choose.
≥ 0.5	Escalate to livepilot-wonder rescue path (not director standard divergence)

Full rules: references/anti-repetition-rules.md.

Honesty Rules (inherited from Wonder Mode)

• Never describe an analytical-only variant as previewable
• Never fabricate distinctness by relabeling the same move
• Fewer than 3 variants is ACCEPTABLE only when, after honestly widening across families AND checking concept packets, fewer real options exist
• On first-pass creative-director calls, actively widen BEFORE accepting a smaller variant set (stuck-rescue is a separate context)

Red Flags — STOP If You Catch Yourself Thinking

Rationalization	Reality
"The user just wants one thing fixed"	Did they name exact parameters? If no, that's creative intent.
"I'll skip the brief, I know what they mean"	The brief pins protected qualities. Skipping them = pattern repetition.
"Three variants are overkill for this"	That's the collapse-to-mode instinct. Generate them anyway, then honestly cull.
"I'll run analyze_mix first to see what's needed"	Critics before divergence pre-converge the answer. Defer.
"The first plan looks good enough"	That's the most-likely completion, by definition. Generate two more.
"Fewer than 3 is fine per Wonder's honesty rule"	True for stuck-rescue. On first-pass, widen FIRST across families, fall back only after.
"Reading anti-preferences is slow"	It's an instant tool. The alternative is generating a rejected pattern.
User said "quickly" / "just do it" / "don't overthink it" / "I'm in a rush"	User pressure framing. The brief is STILL mandatory. Compress PROSE (fewer words around each plan), not STRUCTURE (never skip the brief, Phase 1 reads, 3-family diversity, or preview). Phase 1 reads are instant — "quickly" buys nothing by skipping them.
"User said 'a bit more like X' so novelty should be low"	Correct intuition, but the brief still compiles. "A bit more like X" maps to novelty_budget ≈ 0.45 (see creative-brief-template).
"Two plans that both add sends feel distinct because the sends are different"	Both are mix family with spatial dimension. That is not distinctness. Regenerate from a different family.
"The rhythmic plan has to use a non-arrangement family somehow"	No — rhythmic is a DIMENSION, not a family. Rhythmic plans honestly tag family=arrangement (clip pattern) or family=sound_design (per-hit feel) with dimension_hint="rhythmic". Not a fudge.

What This Skill Does NOT Do

• Does not replace livepilot-wonder, livepilot-arrangement, or any engine skill — it ROUTES to them
• Does not execute arrangement / sound-design / mix tool calls directly
• Does not override user-specified locks ("don't touch X" wins every time)
• Does not fire when livepilot-performance-engine is active (safety wins)
• Does not replace livepilot-core — all Golden Rules still apply

Relationship to Other Skills

Situation	Route to
detect_stuckness > 0.5 or user says "I'm stuck"	livepilot-wonder (rescue path)
Exact mix target ("-14 LUFS")	livepilot-mix-engine minimal-fix mode
Executing a chosen plan (structural)	livepilot-arrangement
Executing a chosen plan (timbral)	livepilot-sound-design-engine
Executing a chosen plan (rhythmic / melodic)	livepilot-notes + livepilot-composition-engine
Post-commit evaluation	livepilot-evaluation with artistic dimensions

References

• references/creative-brief-template.md — YAML schema + filled examples
• references/move-family-diversity-rule.md — how "distinct" is enforced
• references/anti-repetition-rules.md — pre-generation reads + bias rules
• references/the-four-move-rule.md — structural / rhythmic / timbral / spatial coverage