Higgsfield Recall Agent Skill

Higgsfield Recall — Pre-Generation Memory Check

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>" 5

Query 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
• 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

(../../db/ledger/ — 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

• --tags and --reason come 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 --draft for 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|mcsla tags 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-camera

python3 ../../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 to

higgsfield-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 (see higgsfield-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 stats

Empty 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 separation
• higgsfield-soul — Character drift prevention (common recall topic)
• higgsfield-models — Model-specific failure patterns