PocketForge / templates
CLAUDE.md0 starsChanged 7 days agoMITView on GitHub
Pipeline: skill seed → 20 ranked Profitable Pockets → transformation map → competitor benchmark → outline → manuscript → mascot → poses, scenes, animations → cover → layout → PDF + animation pack. Everything runs inside Claude Code: subagents, skills, slash commands, MCP servers, hooks. No Anthropic API key. No custom LLM code. Higgsfield (optional) may be used for cover art, chapter-opening illustrations and promo images only, via node scripts/extras/hf-art.mjs. Never for fake people endorsing the book (no AI-influencer testimonials).…
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- .cursor/rules/pocketforge.mdcCursor rule
- anti-slop-writingSkill
- ebook-structureSkill
- mascot-rigSkill
The file
We split this CLAUDE.md into the rules it gives an agent
11 rules. About 0 are common, 5 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 18
After pockets.md: user picks one pocket.
- Very specific: line 19
After outline.md: user approves title and TOC.
- Very specific: line 20
After assets/character/sheet.png: user approves the mascot.
- Distinctive: line 21
After 3 cover options: user picks one.
- Very specific: line 27
After each phase print exactly one line: ✅ Phase N: <one-line result>.
- Distinctive: line 33
NEVER fabricate stats, testimonials, case studies, names, quotes or credentials.
- Very specific: line 34
All web research goes through the Firecrawl MCP (firecrawl_search, firecrawl_scrape, firecrawl_map, firecrawl_extract).
- Very specific: line 35
Every file in research/ starts with line 1: Source: <URL> | Fetched: <YYYY-MM-DD>.
- Distinctive: line 36
Prefer primary sources (government, associations, peer-reviewed).
- Distinctive: line 85
Ask before: global installs, editing global Claude settings, adding MCP servers, deleting files.
- Distinctive: line 87
No web UI, auth, payments or course generator.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md 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
PocketForge
Pipeline: skill seed → 20 ranked Profitable Pockets → transformation map → competitor benchmark → outline → manuscript → mascot → poses, scenes, animations → cover → layout → PDF + animation pack.
Everything runs inside Claude Code: subagents, skills, slash commands, MCP servers, hooks. No Anthropic API key. No custom LLM code. Higgsfield (optional) may be used for cover art, chapter-opening illustrations and promo images only, via node scripts/extras/hf-art.mjs. Never for fake people endorsing the book (no AI-influencer testimonials). No text or stats baked into generated images.
Entry points
| Command | Does |
/forge "<skill or AUTO>" | Full pipeline, 4 approval stops |
/pockets "<seed>" | Research + 20 ranked pockets only |
/book <slug> | Resume from an approved outline |
/animate <slug> | Rebuild character animations only |
Hard stops (wait for the user, never auto-continue)
- After
pockets.md: user picks one pocket. - After
outline.md: user approves title and TOC. - After
assets/character/sheet.png: user approves the mascot. - After 3 cover options: user picks one.
Record each approval in projects/<slug>/approvals.md (date, stop, choice, user's words). /book and /animate read it to know where to resume.
Phase banner
After each phase print exactly one line: ✅ Phase N: <one-line result>.
Phases: 0 skill check · 1 research · 2 gap mining · 3 pockets (STOP) · 4 transformation map · 5 benchmark · 6 outline (STOP) · 7 manuscript · 8 character sheet (STOP) · 9 poses + scenes · 10 animations · 11 brand + cover (STOP) · 12 layout · 13 export + extras.
Truth rules
- NEVER fabricate stats, testimonials, case studies, names, quotes or credentials. Unsourced claim: cut it.
- All web research goes through the Firecrawl MCP (
firecrawl_search,firecrawl_scrape,firecrawl_map,firecrawl_extract). No other web tool for research. - Every file in
research/starts with line 1:Source: <URL> | Fetched: <YYYY-MM-DD>. - Prefer primary sources (government, associations, peer-reviewed). Mark vendor-blog figures "vendor-reported".
- Health, legal and finance niches get a disclaimer page automatically.
- Benchmarks record patterns only. Never copy competitor text, titles, covers or art.
Writing rules
- Load the
anti-slop-writingandebook-structureskills before writing or critiquing any chapter. - Writers use ONLY their chapter's research files. Every stat gets a
[^n]footnote that resolves to a URL inresearch/. - Mark illustration slots as
[POSE:<name>]and[SCENE:<description>]. Worksheets go in:::worksheet...:::blocks. - Writers run in parallel, max 3 at a time. The critic is always a fresh subagent, never the writer. Max 2 revision loops per chapter, then flag to the user.
- Banned phrases (enforced by
scripts/gates/banned.mjs): delve, unlock, game-changer, landscape, seamless, "in today's fast-paced world", "it's important to note", "in conclusion", "harness the power".
Gates
scripts/gates/ holds citations.mjs, banned.mjs, structure.mjs. The PostToolUse hook in hooks/hooks.json (Claude Code plugin; other agents run the gates by hand) runs all three after every Write/Edit under projects/*/manuscript/ and feeds failures back. A chapter is done only when all three pass AND the critic signs off. Run by hand: node scripts/gates/run-all.mjs <chapter.md>. Tests: npm test.
Images
Allowed only: the code-built mascot, Canva-generated art, Higgsfield art (optional, see above), script-rendered charts/diagrams, royalty-free photos with license logged in assets/credits.md. No decorative images that carry no information. No text baked into generated images.
Mascot
One original mascot per book, a guide character. Load the mascot-rig skill before touching it. Layered SVG rig, flat vector, 3–4 palette colors from brand.md, one line weight. No resemblance to any known character, mascot or brand figure. Every placement points at, reacts to or explains the content beside it. No floating mascots. Consistency check (proportions, colors, line weight) across every pose and animation frame sample before layout.
Render stack (local devDependencies, nothing global):
- Stills:
@resvg/resvg-js(SVG → PNG at 2x, transparent). - Animations: deterministic
seek(t)HTML pages rendered frame by frame with Playwright +ffmpeg-static(scripts/character/render-anim.mjs). Multi-scene clips (framework explainers, promo) use HyperFrames (npx hyperframes) under themotion-doctrineseam law. - Output: MP4 1080p H.264 + GIF ≤ 8 MB.
Canva
Canva is optional and runs through a Canva MCP connector (for example the claude.ai Canva connector). Its tool prefix differs per install; the designer and layout agents inherit all tools, so no prefix needs to be configured. If no Canva tools are available, skip Canva and use the HTML fallback. Any Canva generation outside a /forge or /book run needs the user's OK first. Layout fallback after 2 Canva failures: HTML + print CSS → Playwright PDF (scripts/layout/html-to-pdf.mjs).
Folder conventions
projects/<slug>/
approvals.md research/ pockets.md transformation.md benchmarks.md
outline.md brand.md manuscript/ch01.md … assets/ output/ cost-log.mdbook.json: title, subtitle, author, mascot, blurb, listing and art prompts. Every layout script reads it. Start fromtemplates/project/book.example.json.<slug>: kebab-case from the chosen pocket, max 40 chars.assets/character/{rig.svg, pose-data/*.json, sheet.png, poses/*.png, scenes/*.png, anim/*.mp4, anim/*.gif, qa/}cost-log.md: one line per phase: phase, Firecrawl calls made, Canva calls made, subagents used.- Final deliverables in
output/:ebook.pdf,cover.png,mockup.png,listing-*.png,anim/(MP4 + GIF),README.mdwith a 3-line sales blurb and a price suggested frombenchmarks.md.
Boundaries
- Global skills in
~/.claude/skillsand~/.agents/skillsare read-only. Reference them by name. Never copy or edit them. - Ask before: global installs, editing global Claude settings, adding MCP servers, deleting files.
- Same step fails 3 times: stop, report the shortest decisive error line.
- No web UI, auth, payments or course generator. The web edition of the book is built only if the user asks.
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.

