Aquileo | Improve/skill review optimization by yogesh-tessl · Pull Request #1 · OSideMedia/higgsfield-ai-prompt-skill

yogesh-tessl · 2026-04-14T14:15:55Z

20 sub-skills organised around the MCSLA formula, with a dedicated Seedance preflight linter that catches 11 filter-risk patterns before users burn credits. That's the kind of practical, loss-prevention thinking that separates a real tool from a prompt dump. The fact that seedance_lint.py already exits with code 1 for CI integration tells me you've been thinking about automated quality gates. I'd like to contribute some improvements to a few of the skill files.

I ran your skills through tessl skill review at work and found some targeted improvements. Here's the full before/after:

Changes summary

higgsfield-cinema (+22%) — the biggest win:

Added explicit capability statement to description ("Guides users through professional filmmaking workflows... including creating multi-shot sequences, configuring optical stacks, applying color grading, managing Soul Cast AI actors, and structuring per-scene prompts")
Converted description from YAML folded scalar to quoted string
Consolidated redundant @ Element usage rules (was explained 3 separate times)
Removed duplicate version-mixing warning from output format section
Replaced redundant Prompting Best Practices section with a cross-reference to higgsfield-prompt
Collapsed Cinema Studio 3.0 Resolution Comparison Table and repeated specs into a concise quick-reference list
Removed double --- separator

higgsfield-soul (+18%):

Added explicit capability statement to description ("Creates and manages reusable character profiles (Soul IDs) for consistent facial and stylistic identity...")
Converted description from YAML folded scalar to quoted string
Removed unnecessary explanatory paragraph about why identity drift happens at the model level
Trimmed redundant prose
Condensed mid-example commentary into a single line

higgsfield-model-guide (+10%):

Added explicit capability statement ("Recommends optimal Higgsfield video and image models based on use case, budget, and creative intent...")
Converted description from YAML folded scalar to quoted string
Added inline Quick Decision Flowchart table mapping user needs to recommended models with pricing tiers
Added pricing tier summary (Free / Low / Mid / Premium)

higgsfield-seedance (+1%):

Added concrete actions to description ("Rewrites scene descriptions using professional cinematography language, structures prompts with a six-slot formula...")
Converted description from YAML folded scalar to quoted string

higgsfield-prompt (+0%):

Converted description from YAML folded scalar to quoted string
Consolidated Seedance 2.0 I2V Gate Rule into the earlier I2V section (was duplicated)
Consolidated Seedance 2.0 Narrative Structure Options into the earlier Narrative Structure section (was duplicated)
Trimmed explanatory prose that restated what before/after examples already demonstrate

Honest disclosure. I work at https://github.com/tesslio where we build tooling around skills like these. Not a pitch - just saw room for improvement and wanted to contribute.

I also added a lightweight GitHub Action that auto-reviews any skill.md changed in a PR (includes min permissions, uses a pinned action version, only posts a review comment).

This means that it gives you and your contributors an instant quality signal before you have to review yourself (no signup, no tokens needed).

Want to self-improve your skills? Just point your agent (Claude Code, Codex, etc.) at this Tessl guide (https://docs.tessl.io/evaluate/optimize-a-skill-using-best-practices) and ask it to optimize your skill. Ping me - @yogesh-tessl (https://github.com/yogesh-tessl) - if you hit any snags.

Thanks in advance 🙏

@OSideMedia

Hey @OSideMedia 👋 I ran your skills through `tessl skill review` at work and found some targeted improvements. Here's the full before/after: ![score_card](score_card.png) | Skill | Before | After | Change | |-------|--------|-------|--------| | higgsfield-cinema | 60% | 82% | +22% | | higgsfield-soul | 74% | 92% | +18% | | higgsfield-model-guide | 77% | 87% | +10% | | higgsfield-seedance | 88% | 89% | +1% | | higgsfield-prompt | 92% | 92% | +0% | <details> <summary>Changes summary</summary> **higgsfield-cinema (+22%)** — the biggest win: - Added explicit capability statement to description ("Guides users through professional filmmaking workflows... including creating multi-shot sequences, configuring optical stacks, applying color grading, managing Soul Cast AI actors, and structuring per-scene prompts") - Converted description from YAML folded scalar to quoted string - Consolidated redundant @ Element usage rules (was explained 3 separate times) - Removed duplicate version-mixing warning from output format section - Replaced redundant Prompting Best Practices section with a cross-reference to higgsfield-prompt - Collapsed Cinema Studio 3.0 Resolution Comparison Table and repeated specs into a concise quick-reference list - Removed double --- separator **higgsfield-soul (+18%)**: - Added explicit capability statement to description ("Creates and manages reusable character profiles (Soul IDs) for consistent facial and stylistic identity...") - Converted description from YAML folded scalar to quoted string - Removed unnecessary explanatory paragraph about why identity drift happens at the model level - Trimmed redundant prose - Condensed mid-example commentary into a single line **higgsfield-model-guide (+10%)**: - Added explicit capability statement ("Recommends optimal Higgsfield video and image models based on use case, budget, and creative intent...") - Converted description from YAML folded scalar to quoted string - Added inline Quick Decision Flowchart table mapping user needs to recommended models with pricing tiers - Added pricing tier summary (Free / Low / Mid / Premium) **higgsfield-seedance (+1%)**: - Added concrete actions to description ("Rewrites scene descriptions using professional cinematography language, structures prompts with a six-slot formula...") - Converted description from YAML folded scalar to quoted string **higgsfield-prompt (+0%)**: - Converted description from YAML folded scalar to quoted string - Consolidated Seedance 2.0 I2V Gate Rule into the earlier I2V section (was duplicated) - Consolidated Seedance 2.0 Narrative Structure Options into the earlier Narrative Structure section (was duplicated) - Trimmed explanatory prose that restated what before/after examples already demonstrate </details> Honest disclosure — I work at @tesslio where we build tooling around skills like these. Not a pitch - just saw room for improvement and wanted to contribute. Want to self-improve your skills? Just point your agent (Claude Code, Codex, etc.) at [this Tessl guide](https://docs.tessl.io/evaluate/optimize-a-skill-using-best-practices) and ask it to optimize your skill. Ping me - [@yogesh-tessl](https://github.com/yogesh-tessl) - if you hit any snags. Thanks in advance 🙏

OSideMedia · 2026-04-14T15:10:56Z

Merged the skill improvements — great catches, especially the cinema deduplication. Skipped the workflow for now as we're keeping CI internal, but the content edits are solid. Thanks for the contribution!

…t Block (G21 + G19) — tightened per Decision B self-review Second content sub-phase of v3.7.7 mega-release (third one-time exception). CH-2 part 1: spatial-prompting vocabulary cluster. CHANGES: skills/higgsfield-seedance/SKILL.md (+130 lines, post-tightening) NEW H2: § Frame Coordinate System (G21) Placement: between § Reference Roles and § Working Modes H3 sub-sections: - Qualitative anchors (left/center/right third; upper/lower third; foreground/midground/background) - Percentage notation (x-position / y-position / frame occupancy %) - Pair the two notations (discipline pattern with worked two-character example) - Not a mathematical guarantee (load-bearing caveat per Stage 3 finding; practical recovery via contact-point clauses) NEW H2: § Spatial Layout Block (G19) Placement: immediately after § Frame Coordinate System H3 sub-sections: - What goes in the block (6-field per-subject structure + multi-character cross-subject relationships) - When to use a Spatial Layout Block (3 triggers + when NOT to use) - Block-and-prompt fit (placement before Dynamic Description; non-replacement of six-slot formula slots) Tightening pass applied (Decision B per self-review): Initial draft was 159 lines (244% of Phase 1c LOCK 1 estimate of 40-65). Peter requested honest self-review against three decision options. Self-review classified 6 paragraphs as orientation/restatement padding rather than load-bearing rules: 1. G21 intro double-framing — 9 → 2 lines (-7) 2. G21 qualitative anchors closing "why" — 5 → 0 (-5; lead sentence already explained why; closing was redundant) 3. G21 percentage notation closing — 4 → 2 (-2) 4. G21 pair-the-two-notations closing — 4 → 0 (-4; rule statement above stood on its own) 5. G19 intro "earns its own name" — 7 → 2 (-5) 6. G19 block-and-prompt-fit second paragraph — 6 → 2 (-4) Total -29 lines. All load-bearing rules preserved untouched (NOT-mathematical-guarantee caveat + recovery advice + 6-field block + 3 triggers + multi-character cross-subject relationships + worked example all stay verbatim). Bias check (per Peter's STEP 5 protocol): the initial PR body's "natural explanatory shape cannot be compressed without quality loss" framing conflated two claims — (a) load-bearing rules exist (TRUE) and (b) every paragraph is load-bearing (FALSE). Tightening pass disproved claim (b); 29 lines of orientation scaffolding cut without losing any rule. Readability improved: load-bearing rules surface immediately without padding between them. Locked decisions inherited from Phase 1c: - LOCK 1: 2b-i is G21 + G19 (split from original 2b) - LOCK 5: NOT-mathematical-guarantee caveat present + load-bearing - LOCK 14: vocab.md bracket/arrow deferred to 2f - LOCK 16: G13b deferred to 2b-ii - LOCK 9: no frontmatter bump (cascade deferred to 2h) - No tag operations; no CHANGELOG entry (both deferred to 2h) Catch #14 instance #1 for v3.7.7 — Phase 1c scope-estimate drift: Phase 1c LOCK 1 estimated 2b-i at 40-65 lines. Post-tightening actual: 130 lines (200% of upper bound, down from 244%). Root cause holds: G21 + G19 introduce foundational spatial vocabulary that needs orientation to position relative to existing taxonomies. The split allocation undershot the natural rule density for vocabulary-introducing content. Forward implication: rule-extending sub-phases (2b-ii, 2c-i, 2c-ii, 2d, 2g) ship rules INTO existing taxonomies — less orientation budget needed. Phase 1c estimates for those likely hold. If 2b-ii or 2c-i also overrun by 2x+, revisit Decision C (revise remaining estimates) retroactively. Catch #14 entry deferred to 2h ceremony. Framework innovation candidate (FI5): "Vocabulary-introducing sub-phases may need more orientation budget than rule-extending ones." Log for feedback_release_workflow.md consideration at 2h release ceremony. Cross-reference readiness: - 2b-ii Group F → cites § Frame Coordinate System ✓ - 2c-ii higgsfield-soul Group I → cites both new H2s ✓ - 2d templates (multi-character-anchor.md + top-down-map.md) → cites both new H2s by name ✓ - 2g DISCIPLINE.md Lock-before-generate (Tier 1) → cites pair-the-two-notations discipline ✓ Validation status: - python3 validate.py: ALL CHECKS PASSED (no new files; existing file edit only) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…t Block (G21 + G19) — tightened per Decision B self-review (#27) Second content sub-phase of v3.7.7 mega-release (third one-time exception). CH-2 part 1: spatial-prompting vocabulary cluster. CHANGES: skills/higgsfield-seedance/SKILL.md (+130 lines, post-tightening) NEW H2: § Frame Coordinate System (G21) Placement: between § Reference Roles and § Working Modes H3 sub-sections: - Qualitative anchors (left/center/right third; upper/lower third; foreground/midground/background) - Percentage notation (x-position / y-position / frame occupancy %) - Pair the two notations (discipline pattern with worked two-character example) - Not a mathematical guarantee (load-bearing caveat per Stage 3 finding; practical recovery via contact-point clauses) NEW H2: § Spatial Layout Block (G19) Placement: immediately after § Frame Coordinate System H3 sub-sections: - What goes in the block (6-field per-subject structure + multi-character cross-subject relationships) - When to use a Spatial Layout Block (3 triggers + when NOT to use) - Block-and-prompt fit (placement before Dynamic Description; non-replacement of six-slot formula slots) Tightening pass applied (Decision B per self-review): Initial draft was 159 lines (244% of Phase 1c LOCK 1 estimate of 40-65). Peter requested honest self-review against three decision options. Self-review classified 6 paragraphs as orientation/restatement padding rather than load-bearing rules: 1. G21 intro double-framing — 9 → 2 lines (-7) 2. G21 qualitative anchors closing "why" — 5 → 0 (-5; lead sentence already explained why; closing was redundant) 3. G21 percentage notation closing — 4 → 2 (-2) 4. G21 pair-the-two-notations closing — 4 → 0 (-4; rule statement above stood on its own) 5. G19 intro "earns its own name" — 7 → 2 (-5) 6. G19 block-and-prompt-fit second paragraph — 6 → 2 (-4) Total -29 lines. All load-bearing rules preserved untouched (NOT-mathematical-guarantee caveat + recovery advice + 6-field block + 3 triggers + multi-character cross-subject relationships + worked example all stay verbatim). Bias check (per Peter's STEP 5 protocol): the initial PR body's "natural explanatory shape cannot be compressed without quality loss" framing conflated two claims — (a) load-bearing rules exist (TRUE) and (b) every paragraph is load-bearing (FALSE). Tightening pass disproved claim (b); 29 lines of orientation scaffolding cut without losing any rule. Readability improved: load-bearing rules surface immediately without padding between them. Locked decisions inherited from Phase 1c: - LOCK 1: 2b-i is G21 + G19 (split from original 2b) - LOCK 5: NOT-mathematical-guarantee caveat present + load-bearing - LOCK 14: vocab.md bracket/arrow deferred to 2f - LOCK 16: G13b deferred to 2b-ii - LOCK 9: no frontmatter bump (cascade deferred to 2h) - No tag operations; no CHANGELOG entry (both deferred to 2h) Catch #14 instance #1 for v3.7.7 — Phase 1c scope-estimate drift: Phase 1c LOCK 1 estimated 2b-i at 40-65 lines. Post-tightening actual: 130 lines (200% of upper bound, down from 244%). Root cause holds: G21 + G19 introduce foundational spatial vocabulary that needs orientation to position relative to existing taxonomies. The split allocation undershot the natural rule density for vocabulary-introducing content. Forward implication: rule-extending sub-phases (2b-ii, 2c-i, 2c-ii, 2d, 2g) ship rules INTO existing taxonomies — less orientation budget needed. Phase 1c estimates for those likely hold. If 2b-ii or 2c-i also overrun by 2x+, revisit Decision C (revise remaining estimates) retroactively. Catch #14 entry deferred to 2h ceremony. Framework innovation candidate (FI5): "Vocabulary-introducing sub-phases may need more orientation budget than rule-extending ones." Log for feedback_release_workflow.md consideration at 2h release ceremony. Cross-reference readiness: - 2b-ii Group F → cites § Frame Coordinate System ✓ - 2c-ii higgsfield-soul Group I → cites both new H2s ✓ - 2d templates (multi-character-anchor.md + top-down-map.md) → cites both new H2s by name ✓ - 2g DISCIPLINE.md Lock-before-generate (Tier 1) → cites pair-the-two-notations discipline ✓ Validation status: - python3 validate.py: ALL CHECKS PASSED (no new files; existing file edit only) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…P78 + P85 + P88 + Group B umbrella — tightened per Decision B precedent (#29) CH-2 sub-phase 2c-i. Creates new sibling reference file under skills/higgsfield-seedance/ documenting Seedance render-failure catalog. SKILL.md gets a minimal forward-link in § When the User Is Already in a Failure Loop directing render-failure cases to the new sibling. Files: - skills/higgsfield-seedance/FAILURE-MODES.md (NEW, 293 lines) - skills/higgsfield-seedance/SKILL.md (+7 lines: forward-link block in § When the User Is Already in a Failure Loop) ## Catalog entries (8 sections; 4-field structure each) Group D (Frame-by-frame editing — consolidates P81 + P134 + P135 + P136 from Ep 2 + Ep 3 source): - § FPS drift and frame-by-frame de-duplication - § Frame-level review is mandatory - § Failed-generation salvage Net-new failure-mode entries (P78 + P85 + P88 from Ep 2 source): - § NSFW false-positive (provider-side) - § Keyframe-consistency forces invention - § Physics-state-anchor Group B umbrella (P56 + P52 + P66 from Ep 2 source; cross-refs 2b-i / 2b-ii shipped content): - § Multi-motion camera overload (cross-refs § Single-vs-multi- shot decision shipped in 2b-ii) - § Spatial-awareness failures (cross-refs § Spatial Layout Block + § Frame Coordinate System shipped in 2b-i) Plus structural sections: - § How to use this reference (4-field catalog format definition) - § Self-repair before delivery (7-item pre-delivery checklist — anchors 2g Pre-delivery discipline cross-ref per content-map line 417) - § Cross-references (8 back-links to SKILL.md sections + sibling skills) ## Scope discipline Pre-tightening: 312 lines (142% of LOCK 1 upper bound 220). Post-tightening: 293 lines (133% of upper bound). 19 lines cut per Decision B preemptive discipline. **Catch #14 instance #2 documented.** Different mechanism from instance #1 (vocabulary orientation budget at 2b-i) — instance #2 is catalog-format structural floor at 2c-i. Same catch class (planning-doc drift between LOCK estimate and shipped content). Honest pre-commit classification: - All 8 catalog entries: LOAD-BEARING (symptom + mechanism + counter source-traceable per-entry) - Self-repair section: LOAD-BEARING (anchors 2g cross-ref) - Cross-references: LOAD-BEARING (in-skill + cross-skill) - Frontmatter: LOAD-BEARING (matches MODELS-DEEP-REFERENCE.md precedent) - Padding cut (19 lines): orientation paragraphs + closing meta- commentary + naming-rationale + sibling-explanation + decorative parallels Surfaced Catch #14 instance #2 to Peter at Gate 4 per kickoff prompt mandate (>1.25× threshold = 275 lines). Decision A-tightened approved: apply max tightening, ship at ~292. Final 293 within target band (290-295). ## FI6 framework innovation candidate (logged for 2h) "Catalog-format sub-phases have higher structural floor than rule-extension or vocabulary-introduction sub-phases. Baseline planning estimate at ~240 lines minimum + ~25-30 lines per catalog entry beyond a minimum 5-entry set. Frontmatter + intro + how-to- use + cross-references account for ~60-80 lines of structural floor before any content." Second framework innovation from v3.7.7 execution (FI5 was first — vocabulary-introducing sub-phases need more orientation budget than rule-extending ones). Both ship in 2h feedback_release_workflow.md update. ## validate.py NO change needed. Per LOCK 20: validate.py uses rglob("SKILL.md") and does not check sub-skill sibling files. FAILURE-MODES.md is not validated by name. Same precedent as existing MODELS-DEEP-REFERENCE.md sibling under higgsfield-models/. Cross-references in FAILURE-MODES.md will not be path-checked (only SKILL.md files are scanned for relative path refs). Cross-references FROM SKILL.md TO FAILURE-MODES.md ARE validated; verified passing. ## Frontmatter Per content-map line 293 directive: name + description + user-invocable false + metadata (tags + version 1.0.0 + updated 2026-05-17 + parent: higgsfield). Matches MODELS-DEEP-REFERENCE.md precedent. validate.py PASSES. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…edance/ + templates/text-overlays/, 7 NEW files, NF3-NF9) — Path B + FI7-B refinement (#31) CH-3 sub-phase 2d. Bootstraps templates library with 7 new files across 2 new sub-directories, plus root SKILL.md cross-link infrastructure. Files: - SKILL.md (+21 lines: 3 new tables in § Check Templates for Genre Match + § Shared Resources) - templates/seedance/top-down-map.md (NEW, 73 lines, NF3, G23) - templates/seedance/multi-character-anchor.md (NEW, 96 lines, NF4, G24, with LOCK 15 BAD/GOOD/GREAT) - templates/seedance/single-character-position.md (NEW, 63 lines, NF5, G25) - templates/seedance/worked-example-two-character.md (NEW, 87 lines, NF6, G26) - templates/text-overlays/slogan.md (NEW, 39 lines, NF7, P153a) - templates/text-overlays/subtitle.md (NEW, 39 lines, NF8, P153b) - templates/text-overlays/speech-bubble.md (NEW, 39 lines, NF9, P153c) ## Templates breakdown ### templates/seedance/ — technique templates (NF3-NF6) - **NF3 top-down-map.md** — Claude meta-prompt template for top- down floor plan + paste-ready Seedance blocking note. Includes LOCK 15 BAD/GOOD/GREAT examples. Source: G23 (Hack 17 screenshot Stage 2). - **NF4 multi-character-anchor.md** — Paste-ready Seedance template for multi-character shots with cross-character relationships. Includes LOCK 15 BAD/GOOD/GREAT examples. Source: G24 (Stage 3 seedance-2-pro-director skill body lines 384-405). - **NF5 single-character-position.md** — Simpler companion for single-character shots. Source: G25 (Stage 3 lines 411-427). - **NF6 worked-example-two-character.md** — Concrete end-to-end fill of NF4 structure with named characters Roco + Lulu in neo-noir alley scene. Source: G26 (Stage 3 lines 574-604, paraphrased + extended with cross-character relationships per LOCK 5/Group I). ### templates/text-overlays/ — text-rendering templates (NF7-NF9) - **NF7 slogan.md** — Display text + entrance animation - **NF8 subtitle.md** — Dialogue-synchronized subtitles - **NF9 speech-bubble.md** — Character-attributed dialogue inside frame - Source: P153 (Stage 3 seedance-2-pro-director skill body lines 83-104, three text templates preserved verbatim) ## Scope discipline | Stage | Lines | % of LOCK 1 (275) | % of FI7-adj (340) | |---|---|---|---| | LOCK 1 estimate | 255-275 | — | — | | FI7-adjusted estimate (Gate 1) | 335-340 | — | — | | Pre-tightening | 465 | 169% | 137% | | Post-tightening shipped | 457 | 166% | 134% | **Catch #14 instance #3 documented.** Same catch class as instances #1 (2b-i vocabulary-introducing scope) and #2 (2c-i catalog-format floor), different mechanism: **per-file structural floor for usable template files** (NF7-NF9 each ran 95% over content-map's raw- template-body estimates). Honest pre-commit classification: - NF3 hit FI7 target (73 / 75) - NF4-NF6 (seedance templates): all LOAD-BEARING; template bodies preserved verbatim per LOCK 14 visual-syntax discipline - NF7-NF9 (text-overlay templates): all LOAD-BEARING; field-fills sections are what make 5-line template bodies usable - Root SKILL.md updates: LOAD-BEARING (discoverability) - Padding cut (8 lines): NF4 weakest see-also entries (-5) + NF6 "What this demonstrates" 5→3 bullets compression (-3) ## FI7-B refinement candidate (logged for 2h) "Per-file structural floor for usable template files is ~40-50 lines minimum (title + purpose + template body + field-fills + see-also), independent of template body size. Short templates (P153-class with ~5-line body) need additional usage scaffolding to be standalone-readable; content-map's raw-template-body estimates underproject usable-file floor by ~2×." This is the third refinement to the FI7 framework family from v3.7.7 execution: - FI7: Content-map per-item baselines beat abstract sub-phase estimates (from 2c-ii test) - FI7-A: Gate 1 structural-assumption verification required (from 2d Gate 1 frontmatter check) - FI7-B: Per-file usable-template structural floor (from 2d Gate 4 scope verification) Five framework innovations total now logged for 2h feedback_release_workflow.md update: FI5 + FI6 + FI7 + FI7-A + FI7-B. ## LOCK references - LOCK 1 — sub-phase 2d scope estimate 255-275 - LOCK 11 — NF5 + NF6 naming alignment: single-character-position.md + worked-example-two-character.md (flat naming, NOT templates/seedance/examples/ sub-sub-dir) - LOCK 14 — vocab.md cross-refs forward-declared (resolve at 2f) - LOCK 15 — BAD/GOOD/GREAT inline examples in NF3 + NF4 (anchors 2g DISCIPLINE.md BAD/GOOD/GREAT pattern cross-ref) - LOCK 9 — NO frontmatter cascade for NF3-NF9 (Gate 1 verified: existing templates have no frontmatter); cascade count stays at 12 operations (range collapses from 12-19 to fixed 12) ## Cross-reference resolution **Resolved at this PR:** - All 7 templates cross-link sibling templates within their sub-directory - NF3, NF4, NF5, NF6 cross-link to: - ../../skills/higgsfield-seedance/SKILL.md § Frame Coordinate System + § Spatial Layout Block + § Per-Image Role Convention + § Single-vs-multi-shot decision + § Filter Model (2b-i + 2b-ii) - ../../skills/higgsfield-seedance/FAILURE-MODES.md § Spatial- awareness failures + § Multi-motion camera overload + § Physics- state-anchor (2c-i) - ../../skills/higgsfield-soul/SKILL.md § Character Anchor Block + § Multi-Form State Tracking (2c-ii) - ../../production-benchmarks.md § Per-character iteration anchor (2a; NF6 worked-example only) - Root SKILL.md tables route user requests to new templates **Forward-declared (resolve in later sub-phases):** - vocab.md § Editing Syntax (bracket notation [0-Ns], arrow notation Establishing → phrase → reaction) — ships in 2f per LOCK 14 - 2g DISCIPLINE.md BAD/GOOD/GREAT pattern will cite NF3 + NF4 inline examples (target exists at this PR) validate.py PASSES (NO frontmatter validation for sibling files per rglob("SKILL.md") scope; relative path refs validated only from SKILL.md files). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@image2

…C-tightened per Catch #14 #4) + camera + prompt extensions (P64+P65+P98+P99+Group A+Group E) (#33) CH-2 + CH-5 sub-phase 2f. Cross-cutting consolidation + vocab.md expansion. Path C-tightened approval shipped LOCK 14's 5 audit- discovery entries + kickoff's 3 forward-cross-ref-resolver entries + 5 overlap entries (total 13). Modality-routing + compound-camera- moves vocab entries deferred to v3.7.8+. ## Files - vocab.md (+56 net lines: 13 entries across 4 NEW H2/H3 sections + 5 EXISTING section extensions) - skills/higgsfield-camera/SKILL.md (+5 lines: P64 sequenced-vs- simultaneous combinations + P65 static-pan-vs-glide distinction in § Combining Camera Controls; P98/P99 pull-back from emotional moment row in § Camera-Emotion Sync table) - skills/higgsfield-prompt/SKILL.md (+62 lines: Group A layered emotion states sub-section in § Generic-Emotion Decomposition; Group E prompt-window hygiene sub-section in § Iteration Rule) ## 13 vocab.md entries (Path C-tightened) ### EXISTING section extensions (5 entries) | Entry | Section | |---|---| | P107 scene-physics lighting | § Lighting Vocabulary | | P119 60/30/10 color rule | § Color rules (new sub-header under § Color Grade Language) | | P55 L-cut audio bridge (extends existing) | § Cut & Continuity Vocabulary | | P57 camera-relative-to-previous-shot | § Cut & Continuity Vocabulary | | Reference roles @Image1/@image2 convention | § Image Reference Notations | ### NEW sections (8 entries across 4 new sections) | Entries | NEW section | |---|---| | P140 bracket notation `[0-Ns]` + P141 arrow notation `Establishing → phrase → reaction` | NEW H3 § Editing Syntax (under § Camera Movement Terminology) | | P161 negative space + P162 crossing rule + Coordinate notation | NEW H2 § Composition Vocabulary | | P160 4-layer motion hierarchy | NEW H3 § Motion Hierarchy (under § Camera Movement Terminology) | | P129 script supervising + State lock | NEW H2 § Production Vocabulary | ## Camera + prompt extensions ### camera SKILL.md - P64 sequenced camera moves (pan then push-in) — paragraph added after § Combining Camera Controls "Avoid conflicting moves" callout, distinguishing simultaneous vs sequenced combinations - P65 static pan vs glide — paragraph added in same area, distinguishing camera-rotates-from-fixed-position vs camera-translates-through-space - P98 + P99 pull-back from emotional moment — new row in § Camera-Emotion Sync table for "Emotional breakdown / character needs space" ### prompt SKILL.md - Group A layered emotion states — NEW H3 under § Generic- Emotion Decomposition. Covers composite emotional states (anxious determination, tired tenderness, bitter amusement) + tiny-detail layering (P67 emotion-not-action + P96 tiny- detail emotion-prompting consolidated) - Group E prompt-window hygiene — NEW H3 under § Iteration Rule. Four hygiene patterns: delete obsolete prompt blocks (P46), editor-adds-atop-existing-creates-contradictions (P70), remove stale reference images (P91), prompt-overload sanitize pass (P117) ## Scope discipline | Stage | Lines | % of LOCK 1 (130) | % of Path C-tightened (200) | |---|---|---|---| | LOCK 1 estimate | 80-130 | — | — | | Path C-tightened projection | 130-200 | — | — | | Actual shipped | 123 | 95% | 62% | Under both projections. **FI5-C held strongly:** - 5 entries shipped into existing section anchors (minimal orientation budget) - 8 entries shipped into 4 new H2/H3 sections (each at structural floor; per-entry shared overhead amortized) - All content classified LOAD-BEARING per pre-commit scan; no Decision B tightening needed ## Catch #14 instance #4 documented Planning-doc drift between LOCK 14 (Phase 1c list of 10 entries) and the 2f kickoff list of 10 entries. Mechanism: silent substitution — kickoff list shares only 5 entries with LOCK 14; 5 LOCK 14 items dropped from kickoff; 5 kickoff items added vs LOCK 14. Surfaced at Gate 1 (preemptive); Path C-tightened approved (ship both lists' load-bearing content; defer 2 non-load-bearing items to v3.7.8+). Same catch class as instances #1-#3, distinct mechanism: - #1 (2b-i): vocabulary-introducing orientation budget overrun - #2 (2c-i): catalog-format structural floor overrun - #3 (2d): per-file usable-template floor overrun - #4 (2f): LOCK content overridden by subsequent kickoff scope refinement (silent substitution) ## FI8 framework innovation candidate (logged for 2h) "Phase 2 execution directives that supersede Phase 1c locks must explicitly acknowledge the supersession; silent substitution constitutes Catch #14 surface." Seventh framework innovation now logged for 2h ceremony: FI5 + FI6 + FI7 + FI7-A + FI7-B + FI5-C + FI8. FI8 distinct from FI5-C/FI7 family (which concern estimation methodology); FI8 concerns directive-vs-lock interaction discipline. Different framework dimension. ## Forward-declared cross-ref resolution verification (critical) All 2d template forward-declared cross-refs RESOLVE at this PR: - Bracket notation [0-Ns] → vocab.md § Editing Syntax ✓ - Arrow notation Establishing → phrase → reaction → vocab.md § Editing Syntax ✓ - Crossing rule → vocab.md § Composition Vocabulary ✓ - Coordinate notation → vocab.md § Composition Vocabulary ✓ - Reference roles @Image1/@image2 → vocab.md § Image Reference Notations ✓ - State lock → vocab.md § Production Vocabulary ✓ 2d templates' cross-refs are now resolved repo-wide. ## Sub-phase pattern observation — FI5-C correlation strengthens Anchor-reuse percentage inversely correlates with scope overrun: - 2b-i (new sections): 200% LOCK - 2b-ii (existing taxonomy): 129% LOCK - 2c-i (new file): 133% LOCK - 2c-ii (mostly existing + 1 NEW H2): 96% LOCK - 2d (all new files): 166% LOCK - 2e (all existing anchors): 34% LOCK - 2f (mixed: 5 existing + 8 new across 4 sections): 95% LOCK FI5-C now demonstrated across 7 sub-phases. Strong empirical correlation; possibly underlying mechanism behind FI5/FI6/FI7-B family. ## LOCK references - LOCK 14 — vocab.md additions (5 of original 10 shipped; 5 dropped per Path C-tightened; v3.7.8+ backlog: deferred items reviewed at 2h) - LOCK 11 — camera + prompt extensions per content-map line 288 - LOCK 9 — vocab.md frontmatter cascade at 2h (no separate cascade in 2f; vocab.md has frontmatter per LOCK 9 inventory) validate.py PASSES. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…1 LOCK 4 verbatim reframe + SC2 anti-bombast + 7 new patterns) (#34) CH-5 sub-phase 2g. DISCIPLINE.md expanded from 126 → 246 lines (+120 net). Most procedure-locked sub-phase in v3.7.7: LOCK 4 verbatim paste discipline + LOCK 10 no-redraft + LOCK 12 7-pattern specification all held. ## File - DISCIPLINE.md (+120 net lines: LOCK 4 verbatim Replacement #1 intro replace + 7 new patterns across 3 tiers + SC2 anti-bombast paradox sub-section + LOCK 4 verbatim Replacement #2 NEW § Author-bundle systematization finding) ## LOCK 4 verbatim paste (character-by-character match) ### Replacement #1 — DISCIPLINE.md intro Replaces existing lines 4-12 with the locked language from discovery ledger Section 11. Two paragraphs naming the author-bundle systematization framing + the 5-skills-15-people-14-days-Cannes production anchor. ### Replacement #2 — NEW § Author-bundle systematization finding Per LOCK 4 spec ("Replaces current Convergent-evolution finding section"). Interpretive note: current DISCIPLINE.md has no § Convergent-evolution finding section (Phase 1c-era version likely had it); applying LOCK 10 spirit by adding the locked language as NEW H2 between patterns and § Source attribution. Functionally equivalent placement. LOCK 10 no-redraft discipline held — zero editorial drift; paste matches discovery ledger Section 11 character-by-character. ## 7 new patterns (LOCK 12) ### Tier 1 (+3 patterns — Workflow Discipline) - **Iteration-is-craft** — Cites production-benchmarks.md § Acceptance-rate anchors (2a) + skills/higgsfield-prompt/SKILL.md § Iteration Rule - **Lock-before-generate** — Cites skills/higgsfield-soul/SKILL.md § Character Anchor Block (2c-ii) + seedance/SKILL.md § Frame Coordinate System (2b-i) + § Per-Image Role Convention (2b-ii) - **Falsifiable Success Criteria** — Cites production-benchmarks.md § Falsifiable AI-cinema success criteria (2a) ### Tier 2 (+3 patterns — Output Discipline) - **BAD/GOOD/GREAT** — Cites templates/seedance/top-down-map.md + multi-character-anchor.md BAD/GOOD/GREAT sections (2d per LOCK 15) - **Skill-with-baked-context** — Cites image-models.md § Nano Banana Pro Workflow positioning (2e) + skills/higgsfield-soul/SKILL.md § Two-Tool Refinement Pipeline (2c-ii) - **Anti-Bombast** + **Anti-Bombast Paradox** sub-section (SC2) — Cites vocab.md § Scene-physics lighting (2f) + seedance-2-pro- director skill's anti-empty-words discipline ### Tier 3 (+1 pattern — Architectural Discipline) - **Pre-Delivery Discipline** — Cites skills/higgsfield-seedance/ FAILURE-MODES.md § Self-repair before delivery (2c-i) All 7 patterns follow existing DISCIPLINE.md pattern convention (~10-12 lines each: H3 heading + framing paragraph + When to apply + Demonstrated in cross-ref). ## SC2 anti-bombast paradox sub-section H4 sub-section under #6 Anti-Bombast pattern. Acknowledges the paradox structurally: anti-bombast discipline can itself slide into bombast about restraint. The fix is to demonstrate the rule (BAD vs GOOD rewrite) rather than to praise the rule. The sub-section applies its own discipline recursively (no marketing voice in describing anti-marketing-voice). ## Scope discipline | Stage | Lines | % of LOCK 1 (180) | |---|---|---| | LOCK 1 estimate | 130-180 | — | | Original FI5-C+FI7-B projection (wrong; assumed 25-30 lines/pattern) | 215-255 | 119-141% | | Revised Gate 1 projection (after verifying actual DISCIPLINE.md per-pattern floor ~11 lines) | 95-100 | 53-56% | | Actual shipped | 120 net | 67% | **Under LOCK 1 upper bound.** No Catch #14 instance #5 surfaced. Honest pre-commit classification: all 7 patterns LOAD-BEARING per LOCK 12 + content-map cross-ref priorities; LOCK 4 paste mechanical per LOCK 10; SC2 paradox sub-section LOAD-BEARING for the framework coherence. ## FI7-B refinement candidate (logged for 2h) "Per-file structural floor varies by file convention, not just by file type. Gate 1 must read the target file's existing patterns to verify per-entry floor before applying FI7-B forward-estimation. FAILURE-MODES.md catalog convention (4-field, ~24 lines/entry) and DISCIPLINE.md pattern convention (3-field, ~11 lines/entry) project very differently." This is the second FI7-B refinement (first was FI7-B from 2d's template structural-floor finding). FI7-B framework is becoming nuanced — applies as a forward-estimation tool, requires Gate 1 verification of actual file convention, magnitude varies by file-internal pattern. Combined for 2h: FI7-B applies with file-convention-specific calibration; the calibration step is required at Gate 1 per FI7-A discipline. ## FI5-C correlation extends to 8 sub-phases | Sub-phase | LOCK actual | Anchor reuse | |---|---|---| | 2b-i | 200% | New sections (G21 + G19 new H2s) | | 2d | 166% | All new files (7) | | 2c-i | 133% | New file (FAILURE-MODES.md) | | 2b-ii | 129% | Existing taxonomy extension | | 2c-ii | 96% | Mostly existing + 1 NEW H2 | | 2f | 95% | Mixed: 5 existing + 8 new across 4 sections | | 2g | 67% | All 7 patterns slot under existing Tier H2s | | 2e | 34% | All existing section anchors | Inverse correlation between anchor-reuse and overrun holds across 8 data points. **Strong empirical pattern — FI5-C is the unifying forward-estimation mechanism for v3.7.7.** ## All v3.7.7 cross-corpus cross-references resolve After this merge, every cross-reference threading across CH-1 through CH-5 resolves to a shipped target: - 2g DISCIPLINE.md patterns cite 2a / 2b-i / 2b-ii / 2c-i / 2c-ii / 2d / 2e / 2f content — all RESOLVE - 2d templates cite 2b-i / 2b-ii / 2c-ii / 2c-i content — all RESOLVED at 2f merge - 2c-ii Two-Tool Refinement Pipeline cites 2a production-benchmarks — RESOLVED at 2a merge - 2c-i FAILURE-MODES.md self-repair section anchors 2g cross-ref — RESOLVED at this PR R6 risk (aspirational content / cross-ref resolution) fully mitigated across v3.7.7 Phase 2. ## LOCK references - LOCK 1 — sub-phase 2g scope estimate 130-180 (actual 120 net, under upper bound) - LOCK 4 — SC1 verbatim language pre-locked from discovery ledger Section 11; both Replacements pasted character-by-character - LOCK 10 — no-redraft discipline held; zero editorial drift on LOCK 4 paste - LOCK 12 — 7 new pattern entries (3-3-1 Tier distribution + Anti-bombast paradox subsumed as sub-section under #6 Anti-Bombast pattern) - LOCK 15 — BAD/GOOD/GREAT pattern's "Demonstrated in:" cites NF3 + NF4 templates' BAD/GOOD/GREAT inline examples (anchored at 2d) validate.py PASSES. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

yogesh-tessl added 2 commits April 14, 2026 16:38

ci: add skill-review GitHub Action for automated skill review on PRs

c1b6ff3

OSideMedia closed this Apr 14, 2026

This was referenced May 16, 2026

v3.7.5: backlog closeout — A-arcs + File 08 layer (e) + Layer 0 validator + 2 cross-link callouts #17

Merged

v3.7.7 sub-phase 2b-i — Frame Coordinate System + Spatial Layout Block (G21 + G19) #27

Merged

OSideMedia mentioned this pull request May 17, 2026

feat: v3.7.7 sub-phase 2c-i — FAILURE-MODES.md (NEW file) — Catch #14 instance #2 + FI6 candidate #29

Merged

6 tasks

This was referenced May 17, 2026

feat: v3.7.7 sub-phase 2c-ii — Character Anchor Block (Group I) + Two-Tool Refinement Pipeline (G17) #30

Merged

feat: v3.7.7 sub-phase 2d — Templates library bootstrap (7 NEW files) — Path B + Catch #14 #3 + FI7-B #31

Merged

This was referenced May 17, 2026

feat: v3.7.7 sub-phase 2e — G18 NBP + Group H + Quick Decision Table | chore: markdownlint MD040 fix for 2d templates #32

Merged

feat: v3.7.7 sub-phase 2f — vocab.md 13 entries (Path C-tightened) + camera + prompt extensions — Catch #14 #4 + FI8 #33

Merged

OSideMedia mentioned this pull request May 18, 2026

feat: v3.7.7 sub-phase 2g — DISCIPLINE.md SC1 LOCK 4 verbatim reframe + SC2 anti-bombast + 7 new patterns #34

Merged

6 tasks

This was referenced May 18, 2026

release: v3.7.7 — release ceremony (frontmatter cascade + USER-GUIDE regen + baseline rotation + markdownlint cleanup + CHANGELOG + framework synthesis) #35

Merged

v3.7.13: Marketing Studio + cross-surface workflow #44

Merged

OSideMedia mentioned this pull request Jun 11, 2026

feat: v3.9.0 — improvement briefs 1+2 (specs layer, preflight, tests, evals, hygiene) #54

Merged

OSideMedia mentioned this pull request Jun 23, 2026

feat: v3.15.0 — series finale: Wave C scheduling + item 6 routing telemetry #66

Merged

OSideMedia mentioned this pull request Jul 26, 2026

feat: v3.22.0 — the harvest wave #89

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve/skill review optimization - #1

Improve/skill review optimization#1
yogesh-tessl wants to merge 2 commits into
OSideMedia:mainfrom
yogesh-tessl:improve/skill-review-optimization

yogesh-tessl commented Apr 14, 2026

Uh oh!

OSideMedia commented Apr 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

yogesh-tessl commented Apr 14, 2026

Uh oh!

OSideMedia commented Apr 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants