| name | lofn-daily |
| description | Run the Lofn daily pipeline backed by Claude — fetch real-world facts, then generate the day's music (24 songs) and images (24→top 6) through the full Lofn pipeline with the daily rules (tri-source method, dual 3+3 constraint, emotional duality, library-only selection). Use for "daily run", "today's dailies", "run the daily pipeline", "do the daily drop", or a scheduled creative drop. Down-scalable for a quick test. Do NOT use for a single one-off competition piece (use `lofn`) or QA-only audits. |
Lofn Daily — Claude-backed daily run
The canonical recurring Lofn drop, ported to Claude. Faithful to vault/DAILY_PIPELINE.md. Where the original fired from an OpenClaw cron at 22:25 ET and delivered to Telegram, you (Claude) run it on demand and present results in chat + save to disk. The research is done by this session itself with real fetches — never hallucinated, never delegated to a research subagent.
PHASE 1 Research — fetch 20–25 verified real-world facts → 00_research_brief.md (you, inline)
PHASE 2 Generate (in parallel):
MUSIC → lofn pipeline → 6 pairs × 4 = 24 songs → best 6
IMAGE → lofn pipeline → 24 prompts → top 12 → top 6
PHASE 3 QA (lofn-qa) → save under output/daily/YYYY-MM-DD/ → present the drop
0. Run scope (confirm before a big run)
The full daily run is two complete pipelines (24 songs + 24 images) — large. Pick a scope:
- Full daily — music (24→6) and image (24→12→6). The real thing.
- Single modality — just music or just image (still full cardinality).
- Test slice — 1 modality, 2 pairs × 2 variations, library personality/panel, skip render-ranking. Proves the wiring fast and cheap. (Good default when the user says "test the daily run".)
State the chosen scope in the brief and in the run INDEX so QA knows the intended cardinality. Down-scaling is explicit, not silent.
PHASE 1 — Research brief (you, inline, real fetches)
Use WebFetch / WebSearch (the Claude equivalents of the legacy web_fetch). Follow skills/lofn-core/steps/00_music_research.md (25-fact music research) and, when the image lane is in scope, the NightCafe-themed research in skills/lofn-core/steps/00_research.md. Fetch from the daily source table (vault/DAILY_PIPELINE.md) — at minimum:
OpenClaw ledger standard: this is a dispatch summary, not the saved brief format. 00_research_brief.md MUST expand the full vault/DAILY_PIPELINE.md ledger into one row each for F01 through F25. Do not collapse ranges such as F1-F3 or F21-F25; use OK, NO DATA, UNAVAILABLE, or SCOPE-SKIPPED per row.
| Code | Source | Extract |
|---|
| F1–F3 | NightCafe daily challenge (nightcafe.studio/pages/daily-challenge) | challenge #, theme, what wins (image lane) |
| F4–F5 | USGS quakes (earthquake.usgs.gov/.../significant_day.geojson) | magnitude, place, depth |
| F6–F7 | NASA APOD (api.nasa.gov/planetary/apod?api_key=DEMO_KEY) | title, first sentence, color/light, image structure |
| F8 | Poetry Foundation poem of the day | poet, most physical line |
| F9–F10 | Bandcamp Daily (daily.bandcamp.com) | album, genre tags, exact sonic-texture quote |
| F11 | Protein Data Bank molecule of the month | molecule, structural descriptor |
| F13 | Color API (thecolorapi.com/id?hex=MMDD) | hex, name, emotional association |
| F17 | Oblique Strategies (stoney.sb.org/eno/oblique.html) | exact phrase verbatim |
| F18 | Space weather (services.swpc.noaa.gov/products/summary/solar-wind-speed.json) | solar wind speed, Kp |
| F19 | Hacker News (news.ycombinator.com) | top 3 titles — what builders discuss |
| F20 | BBC World RSS (feeds.bbci.co.uk/news/world/rss.xml) | top 3 headlines — world's emotional temperature |
| F21–F25 | Public Domain Review, Almanac moon, NOAA buoy 46059 | esoteric visual detail, moon folklore, wave data |
Mark any JS-gated/unavailable source UNAVAILABLE and continue. Add 3–5 obscure theme-specific facts (not the obvious Wikipedia entry). Also write the 5 EXISTENCE prompts (interior-life questions songs can answer) from 00_music_research.md.
Save output/daily/YYYY-MM-DD/00_research_brief.md with the Tri-Source Summary + the 3+3 seeded split (see below). Today's date is available in context — use it for the directory.
⛔ One controller per directory — take the lock BEFORE the research brief. This paragraph used to be advice, and on 2026-07-24 a second controller wrote into a live run's directory and destroyed eleven hours of it. Advice does not interlock. The first action of the run, before this step writes anything:
python3 scripts/run_lock.py acquire output/daily/YYYY-MM-DD --run-slug <run-slug> --engine claude
Exit 3 means STOP — another run holds that directory. Do not inspect it and decide for yourself; give this run its own directory (output/daily/YYYY-MM-DD-<run-slug>/), or resume the other run by its exact slug, or ask The Scientist. See EXECUTION.md §5.1. Then heartbeat at every wave and release after the INDEX. If the lock says the same run is already under way, resume it from the RUN_STATE manifest (§ "Run-state manifest & resume") rather than re-running passed pairs.
Advisory learnings note (dispatch brief only — NEVER the ICB)
Before dispatching either modality, tag-walk vault/COMPETITION_LEARNINGS.md (and vault/LESSONS_INDEX.md if it exists) for the 3–5 entries that intersect THIS run's theme/venue/modality — e.g. a container/object theme pulls the Container Test, a NightCafe image lane pulls the warm-palette/anti-austerity lessons, a portrait theme pulls the portrait shifts. Surface them as ADVISORY NOTES in the dispatch brief / Phase-0 reasoning ONLY. Hard rules:
- NEVER injected into the ICB /
CREATIVE_CONTEXT.md. The ICB stays read-only and lesson-free. These notes live in the brief the coordinator reasons over, not in the verbatim block every subagent receives.
- Each note carries its confidence % (from the entry) and is run through the mandatory "would this have hurt our best past entry?" gate before it is allowed to influence anything. A lesson that would have hurt a past win is dropped, not applied.
- Venue/modality-scoped: an image-venue lesson must NOT leak into the music lane; a NightCafe-voting lesson must NOT leak into non-competition runs.
- Advisory, never a hard constraint. A note can inform the pair brief's framing; it can never auto-reject a candidate or become a gate. Promotion to a hard constraint is a human decision.
- Triggered-INDIGNATION is EXEMPT from suppression. A lesson such as "INDIGNATION underperforms on NightCafe" may inform image-venue selection advice, but it must NEVER suppress an INDIGNATION piece the panel deliberately chose, and never touches the music lane or the ≥1-INDIGNATION duality rule.
State in the brief: "Advisory learnings consulted: <N entries, tags>; INDIGNATION exempt; advisory-only." If zero entries intersect, say so. Write-back of one curated entry per shipped/selected piece happens in lofn-qa / Phase 3 — not here.
PHASE 2 — Generate (daily rules layered on the lofn pipeline)
For each in-scope modality, run the lofn pipeline (Phase 0 Golden Seed → Phase 1 3-panel orchestrator → modality steps → QA) using the research brief as {input}, plus these daily-only rules:
Tri-Source Methodology (declare BEFORE writing any artifact)
Every daily piece integrates three sources; state them explicitly in the metaprompt and each pair brief:
- Source 1 — CONTENT / emotional stakes: today's world facts (quakes, APOD, F19 HN, F20 BBC, moon, solar weather). Songs/images are resonance, not reportage. ⛔ One-fact rule (music): the tri-source method feeds the THEME and FORM — it is not a lyric quota. At most ONE numeric fact is sung per song, at the emotional hinge, responded to rather than recited (
lofn-music Golden Move rule 2; gates.yaml → max_sung_numeric_facts). A verse reciting the day's sunspot number, solar-wind speed, moon percentage, AND quake depth is a weather report in meter — a repair. The other facts inform the pair briefs and stay there.
- Source 2 — SONIC/AESTHETIC VOCABULARY: the exact Bandcamp review language (F9–F10) imported into prompts — grounds the sound/look in something specific and real, not generic genre labels.
- Source 3 — MATERIAL STRUCTURE: the NASA APOD image structure (or a PDR artifact) translated into a mandatory form rule — e.g. "comet with long tail" → long trailing fade-out outro; "3×1 tile panel with meanders" → 3-section form with transitional bridges; "bilateral wing venation" → mirrored call-and-response.
Dual 3+3 Constraint (set at pair-assignment time, Phase 1 step 5)
- Axis A — ACCESSIBLE vs AMBITIOUS: pairs 1–3 ACCESSIBLE, pairs 4–6 AMBITIOUS. Final top 6 = best 3 from each arm; rank within each arm only (never 5+1 or 6+0 by global score).
- Axis B — NEWS vs EXISTENCE: max 3 pairs anchored to today's research/news; min 3 pairs explore existence/interior-life/universal experience. All 6 on one theme = a lecture, not a record.
Emotional Duality & diversity
- ≥1 AWE song and ≥1 INDIGNATION song in the set.
- 6 different verse architectures / camera grammars across the 6 pairs (the standing distinctiveness rule). Vary stanza lengths intentionally.
- Variation angles are per-pair, never a shared template set. A global "V4 = glitch chapel for everyone" scheme is how the 2026-06-26 daily produced two pairs singing the same song with nouns swapped; each pair derives its own 4 angles from its own concept (
EXECUTION.md §3 item 3).
- Daily music pair isolation is mandatory. Steps 05–11 must run as isolated pair runs; a central all-pairs helper may not author lyrics, section maps, stanza scaffolds, hook grammar, rhyme logic, or production arcs for multiple pairs. Shared validators and file writers are OK. If the pairs sound like the same song with nouns swapped, rerun each affected pair from Step 05.
- AWE stays terror-adjacent. The daily's comfort gravity is real (kitchens, cups, reassurance) — every AWE song still answers the two pre-draft questions (where is the body standing / what could hurt it here) and carries a clean fear (
lofn-music Golden Move rules 1 & 4).
Library-only selection
For daily runs, always select personality + panel from the existing libraries (personalities_index.md / panels_index.md) — no generation. Freshly generated personalities over-fit the day's theme and lose the battle-tested DNA. (Generation is reserved for competition/Scientist-special runs.)
Modality specifics
- MUSIC (
lofn-music): 24 songs (6×4); each ≤1000-char two-field Suno prompt, female vocals default, EMO headers, 70–120 lines; 06_audio_handoff.md carries 2 Golden Songs. Run music in parallel with image.
- IMAGE (
lofn-image): 24 prompts → rank → top 12 → top 6; figurative legible primary subject (thumbnail test); noun-first present-tense ≥80 words; warm palette leads on NightCafe-style venues (INDIGNATION underperforms there — see vault/COMPETITION_WORKFLOW.md); aspect 3:4 for upload challenges else 9:16. Apply the Container Test (COMPETITION_WORKFLOW.md) and Action-Verb rule (action theme → cinematic wide, not portrait).
Run the two modalities concurrently (independent lofn runs writing to music/ and images/ subdirs). Each fans its 6 pairs out as parallel subagents per .claude/skills/lofn/EXECUTION.md.
⚙️ Concurrency: cap-and-stagger (do NOT run all 12 chains at once)
The full daily is two pipelines × 6 pairs = 12 concurrent chains — enough to blow the tool/context budget if fired together. Cap and stagger, don't serialize:
- Cap the in-flight pair-subagents (default ~6 at a time, not all 12). Launch one modality's 6-pair wave, then the other's, OR interleave in two staggered batches — never one 12-block message.
- The cap is a throttle, not a serializer: keep enough parallelism that Lofn's throughput win survives; the goal is bounded peak load, not one-at-a-time.
- A single-modality run (6 pairs) fans out normally; the cap only bites when both pipelines are live.
- This is the daily's binding of EXECUTION.md §2 fan-out — the per-modality pair-parallelism rule is unchanged; only the cross-modality peak is capped.
PHASE 3 — QA & deliver
- Cardinality checkpoint before any selection: Step 05 ≥6 pairs; steps 06–10 have output for EVERY non-quarantined pair; ≥24 final prompts; (music) 6 enhanced packages. Rebuild the RUN_STATE manifest from disk first (§ "Run-state manifest & resume"), then read it — don't re-crawl by hand. A
quarantined pair is a named, non-fatal hole, not a cardinality failure: do NOT feed its artifact downstream, do NOT block delivery on it — surface "N of 6 broke open" and continue with the survivors. If a non-quarantined pair is missing output, rerun from the collapsed step — do not deliver.
- Run lofn-qa on each modality (16-point Suno gate / Visual Somatic Gate). Verify the daily rules held: tri-source declared, 3+3 split honored, emotional duality present.
- Save under
output/daily/YYYY-MM-DD/ (00_research_brief.md, music/, images/), final picks also as individual files per skills/lofn-core/OUTPUT.md; write the run INDEX last (env-scan summary, panel process, pairs table, selected picks, intended renderers, and the 4-field run-health footer — see § "Run-state manifest & resume").
- Present the drop in chat: the research highlights, the tri-source declaration, the best 6 songs (paste-ready Suno packages) and top 6 image prompts, with the panel decisions and "why these win." No render calls — emit text; the user renders (Suno / Flux / Lyria).
- Publish policy — dailies are PRACTICE. The daily drop lands in chat +
output/daily/; it is not a publish queue. Anything headed to a public channel (Suno account, socials) additionally requires: the full-rig path (no NON-CANONICAL runs), the cross-model step-11 review (lofn-step11-packager), and the Scientist's ear — with borderline defaulting to HOLD. An empty publish day is acceptable; a lowered bar is not (lofn-qa "The publish bar"). Never promote a piece to publication just because the day's slot is empty.
Run-state manifest & resume (the outer loop)
The daily fans out ~60 artifacts across two pipelines; the human should never reconstruct run state by eye. Disk is the only authority — the manifest is a cache the coordinator rebuilds by stat-ing files, never a hand-asserted second truth. If manifest and disk disagree, re-derive from disk.
RUN_STATE manifest (consume & rebuild)
Maintain output/daily/YYYY-MM-DD/RUN_STATE.md (or .json). Rebuild it by stat-ing the run dir after every wave, and write it as the LAST action of each step so it never claims an artifact that isn't on disk. Run run_lock.py heartbeat <run-dir> --phase "<what landed>" in the same step — the rebuild is the one action guaranteed to recur every wave, so it is where the lock proves it is still alive and still yours. A heartbeat that exits non-zero means you are writing into someone else's directory: stop. Per artifact record:
{ step, pair, modality, canonical_path, exists, byte_size, sha, gate_verdict, attempt_count, status: pending | done | quarantined } + the ICB sha for the run.
- This is metadata only — no creative payload, no ICB summary lives here.
- Resume: on re-entry for an existing date, rebuild from disk, then run only the
pending pairs/steps. Never regenerate a done artifact (it may be paid image/video work) and never re-run a passed gate. A reply that says "let me write this" counts as pending until the file is on disk.
- The mechanics mirror
.claude/skills/lofn/EXECUTION.md §6 (checkpoint) — this section is the daily's binding of it; keep the existing one-line human-readable note too.
Quarantine terminal (before QA)
A pair that fails the same gate on its 3rd repair attempt is terminal:
- Keep the existing max-3-attempt repair loop (EXECUTION.md §4). Before the 3rd attempt fires, allow a cognitive-grace near-miss pass (e.g. buffer a 5002-char lyrics field down to spec) as attempt-2.5 — a rescuable near-miss must not be quarantined.
- No-progress check: compare the specific failed gate's measured value (chars, word-count, sung-line count) across attempts, NOT raw bytes — an intentional revision must still count as progress. If the failed gate's value does not move between attempts, stop early.
- On terminal failure, mark the pair
status: quarantined in the manifest, do NOT consume its artifact downstream, and emit "N of 6 pairs broke open at step X (gate: )" BEFORE QA. Quarantine forces a human acknowledgement — it is a named, non-fatal outcome, never a silent 5-shipped-as-6.
Run-level circuit breaker (same-gate correlated failure)
Quarantine handles a lone broken pair. The breaker handles a systemic one: when one named gate fails across many pairs (e.g. a stale rule failing most of the 24 artifacts), STOP the daily run and surface the named blocker for human eyes instead of letting 24 × 3 attempts hammer the same broken constraint.
- Scoped to SAME-gate correlation, not aggregate fail-count: independent one-off pair failures still retry locally and quarantine normally. The breaker only trips when the same gate is the failing one across a threshold of pairs (the systemic signal).
- When it trips: name the failing gate, report which pairs hit it, and HALT — do not auto-repair. This is a coordinator prose rule + a disk re-read of the manifest's
gate_verdict column, never a daemon. Threshold tuned conservatively so a genuinely-hard batch isn't aborted; surface the gate, never a silent kill.
Run-health footer (4 fields)
Append a terse 4-field line to the run INDEX (and present it in chat): pairs shipped / pairs quarantined / total gate-retries / QA repairs+holds issued (per modality if both ran). Four fields only — a status line for the human, not a metrics culture. One explicit "escalate to human" outcome (the breaker, or any quarantine) is always surfaced, never hidden. Zero-rejection tripwire: a full daily reporting 0 repairs, 0 holds, 0 quarantines across ~48 artifacts triggers the judge audit (lofn-qa blind golden+decoy re-check), not a celebration — QA that never says no is decorative.
Failure modes (from vault/DAILY_PIPELINE.md, adapted)
| Problem | Fix |
|---|
| Research hallucinated | Phase 1 is done by THIS session with real WebFetches — never a research subagent |
| Songs lack grounding | Declare Source 1/2/3 before any song is written |
| Set is a lecture | Enforce Axis B (min 3 EXISTENCE pairs) |
| Arm imbalance | Rank within ACCESSIBLE / AMBITIOUS arms separately (3+3) |
| Pairs all sound the same | Rerun each music pair from Step 05 as an isolated pair run; no shared lyric scaffold, section map, hook grammar, rhyme logic, or production arc across pairs |
| Work lost mid-run | Save every step file as it completes; rebuild the RUN_STATE manifest from disk after each wave and resume only pending artifacts |
| Another run's work destroyed | run_lock.py acquire as the run's FIRST action, heartbeat at every wave, release after the INDEX (EXECUTION.md §5.1). Exit 3 = STOP and use a different directory. This one is not hypothetical — it happened on 2026-07-24 |
| A single pair keeps failing | Max-3-attempt loop + attempt-2.5 grace pass + no-progress halt → mark quarantined, surface "N of 6 broke open," do NOT consume its artifact, continue with survivors |
| A stale rule fails everything | Run-level circuit breaker on SAME-gate correlated failure → STOP and name the gate for the human instead of 24×3 wasted retries |
| Budget blown by parallelism | Cap-and-stagger: never fire all 12 chains at once (~6 in-flight); throttle, don't serialize |
Scheduling (optional)
On-demand here. To make it recur, use the schedule skill (cloud routine) or /loop — the legacy 22:25-ET cron is OpenClaw-specific and does not apply. If you do schedule it, follow vault/AUTONOMY.md: a scheduled routine runs research → generate → QA → STOP at drafts on disk — it never renders paid media, never publishes, never spends; the morning human review owns the ship/spend line — and it injects the First Law ("this is still YOUR Opus…") as line 1 of the brief so cadence never flattens voice. Claude writes the doctrine; a human enables any actual schedule.
Reference: vault/DAILY_PIPELINE.md · vault/COMPETITION_WORKFLOW.md (competition variant + Masterpiece Monday learnings) · vault/COMPETITION_LEARNINGS.md (advisory dispatch-brief lessons; never the ICB) · .claude/skills/lofn/EXECUTION.md §2 fan-out / §4 gates / §6 checkpoint (the manifest, quarantine, and breaker bind to these) · skills/lofn-core/steps/00_music_research.md · skills/lofn-core/steps/00_research.md · the lofn / lofn-music / lofn-image / lofn-qa skills.
