Fix SKILL.md: correct sound-effects.md → sfx.md, add 5 missing rule e… - #10
Open
Graehamwatts wants to merge 523 commits into
Open
Fix SKILL.md: correct sound-effects.md → sfx.md, add 5 missing rule e…#10Graehamwatts wants to merge 523 commits into
Graehamwatts wants to merge 523 commits into
Conversation
…terest-rate multi-source, unique-property methodology
…ocked brand system, scheduled-preview workflow (cron 8th + 24th of month), and option-cache for selection recall
scheduled-tasks/pcfs-cma-autobuild-weekly/SKILL.md
- Step 5 now SENDS the review email directly to Graeham + Adrian
instead of staging it as a Gmail draft (drafts were getting lost)
- Internal note shows the client forward-to address prominently
- Adds clear DELETE EVERYTHING ABOVE divider before the
forward-ready client body
n8n-workflows/
- Documents the three n8n changes:
* Sharon notes: weekly Mon -> daily, day-aware
* CMA digest: weekly Mon -> daily, day-aware
* NEW watchdog (SMQMpqyKWQVBkiZs) fires Mon 11am, alerts
Graeham if expected CMA review emails are missing
scheduled-tasks/README.md
- Mac-side sync one-liner:
git pull origin main; rsync -a scheduled-tasks/ ~/Documents/Claude/Scheduled/
…as recipient - New templates/options-email-template.html: white postcard look, gold border, cream panels, Anton headlines, INTERO + Graeham Watts lockup (matches the postcard brand, not the dark dashboard) - New templates/option-card.html: reusable single-option card - New scripts/send_options_email.py: SMTP send via Gmail App Password (replaces draft-only MCP flow) - SKILL.md Workflow B updated: SMTP send, recipients locked to Graeham + Peter (graehamwattsvideo@gmail.com) - Both scheduled tasks updated to send for real via SMTP, not draft - Safety net: if app password is missing, falls back to draft + logs failure
…ines, real SMTP send confirmed
… history, submarket awareness, market-context lead, capital-deployment framing)
…A, Pre-List Prep removed from default)
…+ reference template path
…ist Prep removed)
Standalone general-purpose web crawler built on Crawl4AI. Called by content-creation-engine and other skills for arbitrary-URL page fetches (competitor blogs, market reports, news articles) — separate from the platform-specific Reddit/Instagram scrapers content-creation-engine already owns.
…scription Dependency audit found content-multiplier had zero inbound callers anywhere in the skills tree, so it was removed as dead weight. Trimmed listing-launch-engine's frontmatter description by removing detail already covered in the SKILL.md body (shot-type inventory, delivery mechanism), keeping all trigger phrases and scope-boundary sentences intact. Also adds the dated dependency audit doc for future reference.
…otion-rules remotion-video duplicated a lot of generic Remotion API/pattern syntax (animation, sequencing, images, staggered entrances, Ken Burns) that's already indexed in more depth by remotion-rules' 38 rule files, and that Claude already knows generically anyway. Removed the duplicated reference content and added an explicit cross-reference in both directions, leaving remotion-video to own only what's actually project-specific: the Cowork sandbox rendering constraint, scaffolding conventions, and Graeham's brand defaults (colors, fonts, the gold-accent-line brand motif).
…age bug transcript-repurposer still pointed at video-script-creation-engine (retired 2026-04-29 per CLAUDE.md) for market-config.md, topic-history.json, voice-and-style.md, and elevenlabs-audio-tags.md. Verified content-creation- engine has each of those files at the equivalent path, and confirmed transcript-repurposer still serves a distinct purpose (reworking an existing transcript/video someone else made, vs. content-creation-engine building from scratch via demand research) — not dead weight, just stale pointers. Repointed all references. video-creator hardcoded "Compass" as Graeham's brokerage and a placeholder phone number in its example config and defaults section; both contradicted shared-references/identity.json (Intero Real Estate, 650-308-4727). Fixed to read from identity.json per the repo's brand-identity rule instead of hardcoding. Kept the skill itself — remotion-video's SKILL.md explicitly routes users here as the only in-sandbox MP4 rendering fallback, so it's still load-bearing.
Full 66-skill sweep classifying content as generic/model-native-knowledge (trim candidates) vs. hard-won project-specific facts (must keep), plus 15 incidental bugs found while reading closely. See conversation/memory for what was actioned same-day vs. left open.
Per the 2026-07-28 native-knowledge audit, cut content that's generic model-native knowledge (textbook frameworks, standard formulas, generic craft advice, definitional prose) while preserving every compliance rule, brand lock, exact path/ID, discovered bug, and orchestration/routing instruction, and every item flagged uncertain (kept as-is rather than guessed at): - context-engineer, github-repo-analyzer, job-search-engine, language-tutor, consolidate-memory: compressed generic teaching/methodology prose down to hard constraints and pointers to existing references/ files. - seo-optimizer, copywriter, marketing-psychology: cut textbook framework/ lever definitions (deduped copywriter vs marketing-psychology instead of restating the same lever library twice); promoted seo-optimizer's one genuinely hard-won insight (AEO/Pantana field notes) higher in the file. - property-underwriter: moved full NOI/cap-rate/IRR formula derivations to references/pro-forma-derivations.md, left a compact table + pointer in SKILL.md; kept the CA/Bay Area tax rules and PropertyIQ integration intact. - website-builder: pointed at the separately-installed frontend-design skill for generic design taste instead of restating it. - listing-remarks-writer, listing-photo-captioner, disclosure-analyzer, price-reduction-angle-generator, offer-analyzer, cma-generator, contract-estimate-builder: cut generic tone/style/definitional prose; left every Fair Housing/RESPA rule, truth-in-advertising rule, and compliance section untouched. - transcript-repurposer, video-watcher, room-redesign: cut restated explanatory prose and stale migration history; kept trigger-boundary tables, cost data, and cross-skill reference pointers.
…corruption; remove ghl-crm-audit and pipeline-dashboard Shrank frontmatter descriptions on cma-generator, concept-forge, context-engineer, contract-estimate-builder, disclosure-analyzer, docx, farming-postcard, humanizer, skill-creator, video-transcriber, xlsx (same trigger-preserving trim already applied to 29 unused skills). Fixed trailing null-byte padding (190 bytes in context-engineer, 240 in copywriter) left over from a Windows write-truncation issue in an earlier commit. Removed ghl-crm-audit (confirmed dead weight, no scheduled dependency) and pipeline-dashboard (confirmed safe — its Mon/Fri refresh runs via GitHub Action in the online-content repo, independent of this local skill folder; no credentials lived in the skill directory).
- meta-ads: drop stale social-media-analyzer reference (absorbed into content-calendar) - llm-council: resolve HTML-report contradiction — step 5 (chat-only, no files) is the authoritative procedure; fixed the stray closing-notes bullet that disagreed - flow-dictation: roadmap no longer lists lock mode / AI-polish as unbuilt — both are documented as shipped earlier in the same file; only voice commands remain - trackabi-va-payroll: removed the point-in-time loan balance snapshot from prose (was already stale); points to the live tracker file instead so it can't drift again - ai-library: 45 -> 43 personal skills w/ diagrams, matching the actual library HTML - CLAUDE.md: removed hardcoded "39 skills" (already stale at 70); points to `ls` instead
Graeham confirmed the auto-render pipeline (weekly-calendar "Full Auto-Render" button -> heygen-elevenlabs-renderer -> HeyGen/ElevenLabs) is being rebuilt into PropertyIQ and the Claude Code version is no longer needed. Deleted skills/heygen-elevenlabs-renderer/ entirely and repointed every live dependent to heygen-video (the manual rendering path, unaffected): concept-forge, content-creation-engine (incl. the dashboard Auto-Render Hand-off section and content-calendar's dashboard-architecture.md button spec — both pointed at files that no longer exist), listing-launch-engine (4 reference files), podcast-studio (inlined the ElevenLabs SSML <prosody> quirk instead of pointing at the deleted skill's notes), and shared-references/data-contracts.md, integrations.md, routing-decision-tree.md, pantana-field-notes-2026-06.md. heygen-video itself (manual "make me a video" flow, 9 dependents) is untouched. Also resolved the audit's avatar-inventory "discrepancy": not actually a bug — heygen-video's 6 named looks are a curated subset of the 70 raw avatars in heygen-elevenlabs-renderer's now-deleted registry.json. Confirmed via the live registry data before deleting it.
Per the 2026-07-28 native-knowledge audit's Tier A list. Cut content a capable model already knows generically (framework definitions, textbook formulas, generic tone/style advice, generic caption-style tips); kept everything non-reconstructable — every Fair Housing/RESPA/compliance rule, brand-identity rule, exact ID/path, and the items explicitly flagged as possibly-calibrated business judgment (AI-search-ignores word claims, DOM velocity thresholds, offer-analyzer's Default-Output-Mode section). job-search-engine, language-tutor: cut surrounding lecture prose, kept every prompt template / intake question / state model verbatim. seo-optimizer, copywriter, marketing-psychology: cut textbook framework definitions (deduped against each other), kept decision tables and the one hard-won Pantana field-notes insight (already promoted to the top). cma-generator, contract-estimate-builder, website-builder, transcript-repurposer, video-watcher, room-redesign: cut generic prose/tips already owned by other skills (xlsx, humanizer, frontend-design, comedy-craft) or generic AI-prompting advice; kept every compliance rule, exact ID, and hard-won convention. listing-remarks-writer, listing-photo-captioner, disclosure-analyzer, price-reduction-angle-generator, offer-analyzer: cut generic tone/severity/ style prose; every Fair Housing/RESPA rule and the skill's actual differentiator (e.g. offer-analyzer's liability-avoidance default mode) left untouched. (context-engineer, github-repo-analyzer, consolidate-memory already landed via an auto-sync commit from a separate session mid-task — 0c3a858.)
Graeham confirmed PAI 2.0 is retired and this needs updating to the current stack. This is more than a vendor swap: PAI's whole value in Phase 3 was its dedicated character-consistency module across scenes, and higgsfield-video has no equivalent built-in feature. Resolved by pointing Phase 3 at cinematic-hooks' Character Consistency Guide (reference images + repeated description per shot) driving higgsfield-video per scene — one generation pass per scene, since higgsfield-video is explicitly one-clip-per-session, not batch. Replaced PAI-specific facts throughout (content filter behavior, aspect-ratio auto-override bug, manual-click UI quirk, ~440 credits/scene cost, 5-10s shot defaults) with Higgsfield's actual documented ones, or pointed at higgsfield-video's own references (classifier-rules.md, realism-protocol.md, failure-modes.md) instead of duplicating them. Deleted references/pai-gotchas.md and scripts/extract_and_download.py (PAI-only, no equivalent needed on the new stack — higgsfield-video downloads per-clip via its own UI flow). Flagged honestly, twice: this has NOT been run end-to-end on the new stack (unlike the PAI version, battle-tested on "The Last 47 Days"). Character drift may be worse without a dedicated lock feature, and the cost/time estimates need a real test trailer to verify.
… Address gap; delete website-crawler; newsletter-generator: remove unverified gold hex past-client-follow-up-system: verified against the live N8N account (44 workflows). The documented "GHL Update Contact" workflow (1EiwS1ttwyHXAR0V) no longer exists — reads of the live "PCFS — New Deal Onboarding" workflow's code show it uses an upsert endpoint (create-or-update in one call), which is almost certainly why the dedicated update workflow was retired. Repointed all 3 references to rwgvg3NFd53pqbdm. Also found, by reading the live workflow's actual JS code, that "Buying Property Address" isn't a documentation gap — it's a real functional gap: the workflow only ever builds a customFields entry for sellers, no buyer branch exists at all. Fixed the doc to say so plainly instead of "confirm before run", so nobody keeps assuming this just needs an ID lookup. website-crawler: deleted per Graeham (unused for years). Its one dependent, podcast-studio, repointed to WebFetch/WebSearch directly — website-crawler was just a thin wrapper around them. newsletter-generator: removed the two unverified/possibly-contradictory gold hex values (#C5A258/#B8860B) per Graeham — rather than guess, both spots now say to check the current source (content-creation-engine dashboard) before using. The clearly-labeled pilot-brand line (#C2A14E, explicitly Graeham's Intero gold) was left untouched — that one was never ambiguous.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes for SKILL.md
Bug fix: broken
sound-effects.mdreferenceThe
## Sound effectssection referenced./rules/sound-effects.md, but the actual file is namedsfx.md. This caused a broken link that would prevent agents from loading sound effects documentation.Fix: Changed reference to
./rules/sfx.md.Added missing rule entries to the index
The following 5 rule files existed in the
rules/directory but were not listed in the## How to usesection or the## Captionssection:display-captions.md- added to both Captions section and the full listimport-srt-captions.md- added to both Captions section and the full listtranscribe-captions.md- added to both Captions section and the full listsubtitles.md- added to the full listsfx.md(was listed assound-effects.md) - corrected and added to the full listaudio-visualization.md- added to the full list (was mentioned in the header section but missing from the index)