higgsfield-recall
Skill788 starsChanged 51 days agoMITView on GitHub
Use this skill AUTOMATICALLY before writing any Higgsfield prompt. Query the memory databases for relevant past failures and pre-apply known fixes before the user even hits generate. Triggers include: any request to write a Higgsfield prompt, any use of the higgsfield-prompt skill, any mention of generating a video or image on Higgsfield, any MCSLA prompt construction. This skill should run SILENTLY in the background — don't announce it, just apply what's known. If the databases are empty, skip silently and proceed with normal prompt generation.
In this repository36 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- higgsfieldSkill
- higgsfield-3dSkill
- higgsfield-actingSkill
- higgsfield-appsSkill
- higgsfield-assistSkill
- higgsfield-audioSkill
- higgsfield-cameraSkill
- higgsfield-canvasSkill
- higgsfield-character-designSkill
- higgsfield-cinemaSkill
- higgsfield-content-factorySkill
- higgsfield-facsSkill
- higgsfield-gpt-image-2Skill
- higgsfield-image-shotsSkill
- higgsfield-marketing-studioSkill
- higgsfield-mixed-mediaSkill
- higgsfield-modelsSkill
- higgsfield-moodboardSkill
- higgsfield-motion-designSkill
- higgsfield-motionSkill
- higgsfield-pipelineSkill
- higgsfield-promptSkill
- higgsfield-recipesSkill
- higgsfield-repurposeSkill
- higgsfield-scene-engineSkill
- higgsfield-seedance-2-5Skill
- higgsfield-seedanceSkill
- higgsfield-seedance-vfxSkill
- higgsfield-shotlist-directorSkill
- higgsfield-soulSkill
- higgsfield-stackSkill
- higgsfield-styleSkill
- higgsfield-troubleshootSkill
- higgsfield-vibe-motionSkill
- higgsfield-workspacesSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
19 rules. About 0 are common, 16 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈16
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 48
Before writing any Higgsfield prompt, query both memory databases to find relevant past failures.
- Distinctive: line 64
Run a recall check whenever:
- Distinctive: line 79
Before querying, pull the key semantic terms from what the user wants:
- Distinctive: line 103
Use 3–6 of the most specific nouns from the prompt
- Distinctive: line 104
Run separate queries for the subject, action, and style if needed
- Distinctive: line 120
Relevance threshold: Only act on entries with a relevance score > 0 from the query.
- Distinctive: line 128
Remove or substitute the blocked terms before presenting the prompt
- Distinctive: line 129
If a substitution was confirmed to work, use it directly — except where it breaks a hard engine rule.
- Distinctive: line 135
Do not tell the user "I removed X because it was blocked before" unless they ask — just present the clean prompt
- Distinctive: line 139
Use the confirmed improved prompt structure as the base
- Distinctive: line 147
Only mention the recall results if:
- Distinctive: line 162
If nothing relevant found: proceed silently, no mention of the recall check.
- Distinctive: line 187
Before finalizing any prompt, check:
- Very specific: line 200
Every generation attempt belongs in the generation ledger (ledger/ in the memory folder, ../../SKILL.md § Working Folders — kept AND …
- Distinctive: line 209
If the verdict and reason are already clear from what they said, ask nothing — log it directly.
- Distinctive: line 212
If they don't answer, drop it.
- Distinctive: line 213
Write the row yourself:
- Very specific: line 224
Add --draft for 480p exploration rolls (excluded from headline ratios).
- Very specific: line 250
After a few logged rows, python3 ../../scripts/higgsfield_memory.py ratio <project> prints a per-shot-tag verdict that decides …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Higgsfield Recall — Pre-Generation Memory Check
<!-- CONTENTS — generated by scripts/build_index.py; do not hand-edit -->Contents
- Purpose
- When to Run
- Recall Workflow
- Step 1: Extract search terms from the prompt intent
- Step 2: Query both databases
- Step 3: Evaluate relevance
- Step 4: Apply findings silently
- Step 5: Surface findings only when material
- Manual Recall (User-Initiated)
- Pre-Generation Checklist (run mentally before every prompt)
- Log the Generation Result — One Question, One Command
- Optional: log the routing (usage telemetry)
- Read the verdict before re-rolling
- Database Status Check
- Related skills
<!-- /CONTENTS -->Purpose
Before writing any Higgsfield prompt, query both memory databases to find relevant past failures. Apply known fixes silently — the user should never have to remember what broke before. The system remembers for them.
This skill runs automatically as part of any Higgsfield prompt generation. It does not interrupt the workflow unless it finds something relevant.
Bootstrap status: The databases ship with seed entries covering the most common failure patterns (character drift, VHS style ignored, I2V static output, camera conflicts, lip-sync desync, content filter blocks for real persons and IPs). These grow automatically as the user logs new failures.
When to Run
Run a recall check whenever:
- Writing or improving a Higgsfield prompt (any type)
- The user mentions a topic, character, action, or style that could match past failures
- The prompt contains terms that historically triggered content filters
- The model being selected has previously produced poor results for this type of shot
Do NOT announce running the recall check. Just run it, apply what's relevant, and proceed. Only surface findings when they directly change the prompt.
Recall Workflow
Step 1: Extract search terms from the prompt intent
Before querying, pull the key semantic terms from what the user wants:
Extract:
- Subject/character (person type, appearance)
- Action (what they're doing)
- Location/environment
- Style (visual style, model, camera)
- Topic (the general category: "car chase", "product shot", "horror scene")Step 2: Query both databases
# Check for relevant filter blocks:
python3 scripts/higgsfield_memory.py query-filter "<key terms from prompt>" 5
# Check for relevant quality failures:
python3 scripts/higgsfield_memory.py query-quality "<key terms from prompt>" 5Query strategy:
- Use 3–6 of the most specific nouns from the prompt
- Run separate queries for the subject, action, and style if needed
- Prioritize entries with
fix_confirmed: true— these are proven solutions
Step 3: Evaluate relevance
For each result returned, assess:
| Question | If yes → |
| Does this entry's topic/category directly overlap with this prompt? | Apply the known fix |
| Is a blocked term present in my draft prompt? | Remove/substitute it now |
| Did this model fail on this type of shot before? | Consider switching models |
| Is there a confirmed improved prompt for this scenario? | Use it as the base |
Relevance threshold: Only act on entries with a relevance score > 0 from the query. Ignore entries that only match on generic words.
Step 4: Apply findings silently
For filter block matches:
- Remove or substitute the blocked terms before presenting the prompt
- If a substitution was confirmed to work, use it directly — except where it breaks a
hard engine rule. A stored substitution that describes a character by age (the
real-person entry says "age range", and its example names one) loses to
../higgsfield-seedance/ENGINE-RULES.mdrule 1: keep the archetype, drop the age, and describe by role, build and visible markers. The memory record is data and is not rewritten; the rule is applied when the substitution is used. - Do not tell the user "I removed X because it was blocked before" unless they ask — just present the clean prompt
For quality failure matches:
- Use the confirmed improved prompt structure as the base
- Apply the specific fix that worked (e.g. explicit artifact description for VHS)
- Adjust the model if a better one was identified for this scenario
Step 5: Surface findings only when material
Only mention the recall results if:
- A significant change was made to avoid a known filter block
- A model switch is recommended based on past failures
- The recall found a directly relevant confirmed fix that substantially changes the prompt
How to surface findings (when needed):
"⚠️ Filter note: Previous attempts with [term] were blocked on [date].
Using '[substitution]' instead — this was confirmed to pass."
"📋 Quality note: [Model] produced [failure type] for this scenario before.
Switching to [better model] based on past results."If nothing relevant found: proceed silently, no mention of the recall check.
Manual Recall (User-Initiated)
The user can also request a recall check directly:
"What do we know about [topic] failing?"
"Has [model] had issues with [scenario] before?"
"What got blocked when we tried [type of content]?"
"What's our substitution for [blocked term]?"For these queries, surface the full relevant entries with:
- The original failure
- The substitution or fix that was tried
- Whether it was confirmed to work
- The date it was logged
Pre-Generation Checklist (run mentally before every prompt)
Before finalizing any prompt, check:
- Named real person in prompt? → Check filter-memory for real-person blocks
- Weapon, drug, or violence language? → Check filter-memory for violence/substance blocks
- Brand or IP name? → Check filter-memory for brand-ip blocks
- Using a model that has failed for this scenario type? → Check quality-memory
- Using VFX/style keywords that were previously ignored? → Check quality-memory
- Character consistency required? → Check quality-memory for character-drift entries
Log the Generation Result — One Question, One Command
Every generation attempt belongs in the generation ledger
(ledger/ in the memory folder, ../../SKILL.md § Working Folders — kept AND rejected; the denominator is what makes
takes-per-kept ratios possible). The write path is agent-side and obeys the
5-second rule: at most one short question, then the agent runs one
command. The human never formats JSON, never fills a form.
When the user reports a generation result (pastes a link, says "that one worked", "trash", "the face drifted again"):
- If the verdict and reason are already clear from what they said, ask nothing — log it directly.
- Otherwise ask exactly one question: "keep or reject — what failed?" If they don't answer, drop it. Never ask twice, never nag.
- Write the row yourself:
python3 ../../scripts/higgsfield_memory.py log-gen <project> \
--model seedance_2_0 --tags dialogue-cu,two-char \
--outcome rejected --reason extra-cuts --credits 160--tagsand--reasoncome from the controlled vocabularies in../../db/ledger/README.md— map the user's words to the nearest vocab value ("face drifted"→identity-drift); never invent new values.- Add
--draftfor 480p exploration rolls (excluded from headline ratios). - Wrong verdict logged?
python3 ../../scripts/higgsfield_memory.py amend-gen <id> outcome=kept— corrections are superseding rows, history stays. - Project name: the user's production name if one is established in the
conversation, else
default. - Logging
--method quick|mcslatags the row for the framework-lift A/B (ab <project> --tag <shot_tag>); omit it to leave the row unlabeled and out of the comparison — never guess a method.
Optional: log the routing (usage telemetry)
HARD RULE #1 already makes you name the sub-skills you routed to on the first line of every response. When a production is tracking which skills actually earn their keep, persist that declaration:
python3 ../../scripts/higgsfield_memory.py log-route --skills higgsfield-prompt,higgsfield-camerapython3 ../../scripts/higgsfield_memory.py routing then ranks sub-skills by opens and
lists the never-opened long tail. This is instrumentation, not a verdict —
it makes "which skills are load-bearing, which to prune" answerable from data
once enough requests accumulate; a small sample is not evidence a skill is dead.
Read the verdict before re-rolling
After a few logged rows, python3 ../../scripts/higgsfield_memory.py ratio <project>
prints a per-shot-tag verdict that decides iterate-vs-batch:
iterate(structural-dominant) → the prompt is wrong; hand off tohiggsfield-prompt§ The Iteration Rule (one variable at a time).batch+sel(stochastic-dominant) → the prompt is right; stop re-rolling one at a time — lock it, roll a batch, cull (seehiggsfield-prompt§ Batch-and-Select).low-n→ fewer than five rows; don't trust the split, call it by eye.
A ⚠ plausibility line means a tag is beating its planning default by a wide
margin — either real lift or under-logged failures; surface it, let the
user decide. The verdict is only as good as the reject_reason labels, so map
the user's words to vocab honestly — and when the rejected output is in hand,
classify it from the frame instead of from memory (higgsfield-troubleshoot §
Vision-Grounded Diagnosis logs a --vision-reason alongside the human verdict,
advisory until the agreement command proves it).
Database Status Check
To see current knowledge base size:
python3 scripts/higgsfield_memory.py statsEmpty databases = no recall benefit yet. Start logging failures with higgsfield-troubleshoot
and the recall system gets smarter with every entry.
Negative constraints: The recall system complements
../shared/negative-constraints.md. The shared file covers universal prevention rules; this recall system covers user-specific past failures and confirmed fixes.
Related skills
higgsfield-troubleshoot— Diagnose and fix specific failures (feeds recall DB)higgsfield-prompt— MCSLA formula, Identity/Motion separationhiggsfield-soul— Character drift prevention (common recall topic)higgsfield-models— Model-specific failure patterns
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

