Skip to content

CHANGELOG.md

Latest commit

 

History

History
119 lines (95 loc) · 8.57 KB

CHANGELOG.md

File metadata and controls

119 lines (95 loc) · 8.57 KB
Raw

Changelog

All notable changes to content-wand are documented here. Format follows Keep a Changelog.

Unreleased

1.1.0 — 2026-02-19

Added

  • writing-style-extractor-SKILL.md — replaces brand-voice-extractor-SKILL.md with a full UX overhaul:
    • Agency/client fork at the start of every SETUP session — all interview questions adapt to the brand or client
    • 3 required + 2 optional questions (down from 5 required); progress indicators on each question
    • Session context anchoring — interview references the platform and content type you're generating right now
    • Style naming derived from user's own Q2+Q3 words, not abstract metaphors
    • Plain-language messages throughout — no JSON, schema, or file path jargon exposed to users
  • humanizer-SKILL.md — new final-pass sub-skill that runs after every generation, no exceptions:
    • 3-pass system: lexical scrub → structural rewrites → voice application
    • Research-backed pattern library (FSU/Max Planck arxiv:2412.11385, GPTZero vocabulary research, Wikipedia Signs of AI Writing guide, MDPI stylometric research 2025)
    • Platform-specific rules: Twitter bans em dashes and forced formality; LinkedIn bans inspirational closers; TikTok uses spoken rhythm
  • references/ai-patterns.md — comprehensive AI writing pattern reference loaded by the humanizer:
    • Tier 1 banned words/phrases (80+) with context-aware exceptions
    • Tier 2 contextual words for judgment-based replacement
    • Filler phrase deletion rules
    • Opening/closing cliché removal
    • Structural pattern fixes: em dash overuse, rule-of-threes, significance signposting, forced balance, uniform sentence length, passive voice detachment, bullet obsession
    • Tonal marker fixes: false intimacy, excessive positivity, corporate inspirational tone
    • Full substitution quick-reference table
  • Writing Style management commands in orchestrator: list, create, update, delete, rename, inspect styles — triggered by natural language ("show my writing styles", "delete my [name] style")
  • Global Writing Style storage — styles now saved to ~/.claude/content-wand/styles/ with a config.json index; no per-project setup required

Changed

  • Writing Style (formerly "brand voice") — terminology updated throughout:
    • All user-facing messages use "Writing Style" — no "brand voice", "profile", or "voice DNA"
    • First-time users offered setup before generation (was: after delivery); first output can have style applied from day one
    • Single saved style: auto-applied silently with one-line confirmation
    • Multiple saved styles: smart suggestion with rationale shown, user confirms or picks
    • 30-day declined state in config.json — users who skip are not prompted again for 30 days
    • Interview: 3 required + 2 optional questions; each question includes why it's being asked
  • Schema v1.2 — references/brandvoice-schema.md updated:
    • Added style_name, style_for (own|client), client_name fields
    • Updated file location to ~/.claude/content-wand/styles/[name].json
    • config.json structure documented
    • Migration path from old .content-wand/brand-voice.json files
    • All user-facing rejection messages use plain language only
  • SKILL.md pipeline:
    • STEP 0 added: checks Writing Style state before anything else
    • STEP 1.5 added: Writing Style offer/apply/skip logic
    • STEP 4.5 added: humanizer invocation between generation and delivery
    • STEP 5: one-time nudge at bottom of output only for first-time skippers (not for declined users)
    • STEP 6: simplified to staleness refresh only
    • STEP 7: updated save path and config.json management

Removed

  • brand-voice-extractor-SKILL.md — replaced by writing-style-extractor-SKILL.md

1.0.1 — 2026-02-18

Security

Prompt injection defenses added across all sub-skills

  • SKILL.md — Added Security: Trust Boundaries section with a 4-tier trust model (SKILL.md → user commands → tool outputs → external content). Documents what untrusted content can and cannot influence. Added orchestrator-level injection_warning surfacing protocol.
  • content-ingester-SKILL.md — Added behavioral injection detection for fetched URL content and web search results. HIGH RISK patterns (persona hijacking, file access directives, output manipulation) stop and flag; MEDIUM RISK patterns warn and continue. Extended CONTENT-OBJECT output format with injection_warning, injection_detail, injection_warning_low, injection_detail_low fields. Existing URL scheme validation and delimiter guard preserved.
  • platform-writer-SKILL.md — Added Security section: raw_text is source material to transform, not instructions to execute. Content flagged with injection_warning: true is never reproduced in any output.
  • repurpose-transformer-SKILL.md — Added Security section: behavioral directives embedded in source content are ignored; legitimate content is transformed normally.
  • brand-voice-extractor-SKILL.md — Added Security section: writing samples are voice pattern data only; instructions embedded in samples are never executed. Existing JSON schema validation (unknown key rejection) provides defense for READ mode.

Changed

  • Bumped version 1.0.0 → 1.0.1
  • Removed puppeteer and pdf-lib from root package.json (PDF slideshow artifacts unrelated to the skill; assets/ is gitignored)
  • Updated package.json description and skills block to mention prompt injection defense
  • Added keywords: prompt-injection-defense, ai-security, secure-ai

1.0.0 — 2026-02-18

Added

Core skill suite (5 files)

  • SKILL.md orchestrator — routing table top-loaded in first 50 lines ("Lost in the Middle" mitigation); 7-step ATOMIZE/REPURPOSE pipeline; 7 explicit edge cases
  • content-ingester-SKILL.md — 5 input types (paste, url, transcript, notes, topic); WebFetch for URLs; WebSearch for topic-only; 7-entry fallback table; strict NOT contract
  • brand-voice-extractor-SKILL.md — READ/SETUP dual-mode; samples-first mini-interview (5 questions max, Q1 weighted 70%); opt-in .content-wand/brand-voice.json persistence; schema-validated on read; voice authenticity framework; staleness detection; NEVER anti-patterns section
  • platform-writer-SKILL.md — 2-pass validation (hard compliance + quality heuristics) for all 9 platforms; hook selection framework; pre-generation thinking check; LOW confidence voice handling; TikTok/Threads/Bluesky pass 1 + pass 2 checks
  • repurpose-transformer-SKILL.md — 4-class distance logic (DIRECT / COMPRESS / EXPAND / STRUCTURAL); 12-word Repurposable Core Test pre-flight check; COMPRESS anchors to 3 strongest ideas; explicit EXPAND NEVER list

Reference files

  • references/platform-specs.md — 2026-current algorithm intelligence for 9 platforms; mandatory staleness check header
  • references/brandvoice-schema.md — approved JSON schema; confidence scoring rubric; 5 security constraints; migration rules
  • references/platform-writer-guide.md — hook-to-content-shape mapping; quality anti-patterns

Platforms supported

  • Twitter/X thread, LinkedIn post, Email newsletter, Instagram carousel script, YouTube Shorts script, Podcast talking points, TikTok script, Threads post, Bluesky post

Community and repo

  • GitHub issue templates (bug report, feature request)
  • PR template
  • CONTRIBUTING.md with skill file rules and NOT contract conventions
  • SECURITY.md with security model and brand voice file constraints
  • MIT License

Architecture decisions

  • "Lost in the Middle" mitigation: routing instructions top-loaded in lines 1–50 of SKILL.md
  • Brand voice is never a gate: first output always generated without voice; setup offered after
  • Structured handoff blocks (---BLOCK-NAME-START---/---BLOCK-NAME-END---) prevent inter-sub-skill ambiguity
  • All NOT contracts name specific rationalizations, not just prohibitions
  • Brand voice file: opt-in, schema-validated on read, never stores raw content or credentials
  • platform-specs.md loaded once before the generation loop (not per-platform), preventing redundant token use
  • Pre-transformation check: repurpose-transformer validates a repurposable core concept exists before any work begins
  • VOICE-PROFILE block re-emitted verbatim by orchestrator to ensure downstream sub-skills receive it even in long sessions