v0.16.0

What changed

Adverbs earn their place: one shared word-level rule for every co-founder and coach.

Added

• shared/persona/adverb-rules.md. A single snippet holding the "Adverbs earn their place" rule, included by all seven advising skills (Jack, Maya, Priya, Dan, team, pitch-deck-coach, startup-application-coach). It covers:

  • No empty intensifiers. Very, really, quite, truly, incredibly, extremely, absolutely, genuinely, literally, honestly, simply, and "just" when it means "merely". They carry no information and make a claim land softer, not harder. When a line needs more force, the fix is a stronger noun or a concrete number.
  • No weak verb propped up by a manner adverb. "Ran quickly" is "sprinted". "Grew quickly" is "doubled" when the number is there to back it. Every -ly word next to a verb gets checked against a single stronger verb.
  • What stays. Adverbs of time, place and frequency (tomorrow, twice, outside, always, still) and linking adverbs (however, therefore, instead) are never cut for being adverbs. A manner adverb stays when it changes the meaning and no single verb covers it.
  • What is never touched. Fixed text, transcripts, quotes from customers or investors, and copy the founder writes themselves, unless an edit is asked for.
  • Non-English copy. The principle holds in any language, but the word list and the -ly test are English-specific.

  The rule applies to conversation and artifacts alike, and runs alongside the humanizer pass rather than replacing it. The humanizer skill is unchanged: its vendored catalog covers hedges and filler phrases, and this covers the word-level habit it does not.

Changed

• The humanizer proof line now covers adverbs. shared/persona/humanizer-steps.md (used by the four personas) and the inline copies in both coaches ask for intensifiers and weak verb plus adverb pairs in the same one-line "show your work" note, so a draft still reports one line, not two.
• hooks/humanizer-slop-check flags intensifier pile-ups. The optional, opt-in Claude Code hook now scans new text for empty intensifiers and reports them as one kind of tell, or speaks up on its own at four or more in a single write. "just" and "simply" are deliberately left off the hook's list: both appear legitimately in docs and code comments too often to flag without noise. The skill rule still covers them. Claude Code only; nothing changes on Claude.ai.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.15.1

What changed

Changed

• Credits in README.md. Updated Charles Hope's figures: 700+ founders (was 650), 17+ startup programs globally (was 14+), and added 26+ years of experience, including 14+ years in and around startups. Docs only, no skill behavior change.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.15.0

What changed

One shared style rule for every co-founder voice: simple plain language, concise, organised.

Added

• shared/persona/talk-rules.md. A single snippet holding the four rules every co-founder voice follows. Jack, Maya, Priya, Dan, and the team skill all include it, so how the team communicates is now edited in one place.
  • Simple, plain language, in whatever language the founder is using. Short words, everyday phrasing, no jargon unless it earns its place, and any unavoidable term explained the first time it appears.
  • Concise. The shortest answer that still says why: the position, the one or two reasons behind it, then stop.
  • Organised. Answer first, reasons after. One idea per paragraph, related points together. A list only when there are genuinely several parallel items, and never headings or bullets around a two-sentence answer. This is deliberately about organising the thinking, not adding formatting.
  • No em dashes, unchanged from before.
  • Plus the scope line: concise applies to conversation, while plain language and organisation apply to artifacts too. Artifacts still run as long as they genuinely need to.

Changed

• The four personas share the style rules instead of restating them. shared/persona/talk-brief.md (em dashes and brevity only) is replaced by talk-rules.md. Each persona's "How you talk" section drops its own duplicated "clear, simple language" and artifact-scope bullets and leads with the shared block. Dan keeps a persona-specific line about explaining dense fundraising vocabulary (MFN, pro-rata, liquidation preference) on first use.
• The team skill now follows the same rules. Its "Language and style" section includes talk-rules.md instead of the one-line summary it carried before. Multi-voice replies are where verbosity does the most damage.
• CLAUDE.md documents the shared style rules as the fifth cross-skill contract.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.14.0

What changed

The humanizer stops being a fork. Its pattern catalog now comes straight from the upstream project and re-syncs itself, so upgrading the bundle also gets you the newest patterns.

Added

• Vendored upstream humanizer catalog. humanizer/references/upstream-patterns.md is blader's humanizer SKILL.md, MIT licensed, copied in unchanged. It replaces the hand-written 28-pattern catalog we used to maintain, which had fallen behind upstream. Users get 35 patterns instead of 28, plus upstream's "what not to flag" guidance, which the old catalog lacked. New patterns cover repeated sentence openings, writing about the previous version, forced punchlines, formulaic sayings, fake-candid openings, answering objections nobody raised, and rejecting fake alternatives.
• sync-humanizer. A plain-bash script at the repo root that pulls upstream's latest tagged release and rewrites the vendored file. bash ./sync-humanizer --check reports without writing. It refuses to vendor anything containing a build directive.
• Weekly upstream sync workflow. .github/workflows/sync-humanizer.yml runs the sync every Monday, rebuilds, validates every skill against the Agent Skills spec, and opens a pull request when upstream has published a new version. Nothing merges on its own.

Changed

• humanizer/SKILL.md is now a thin layer, not a fork. It keeps what is ours (the workflow, output sizing, the English/non-English scope rule, the voice and personality guidance) and hands the pattern list to the vendored catalog. Bumped to 4.0.0 in its own informational version.
• Em dashes are now an absolute rule, not a density judgment. The old fork said one em dash in a long piece was fine. Upstream's catalog bans them outright unless the writer's own sample uses them, and this repo's house style bans them too. humanizer/SKILL.md now follows the catalog, and names em/en dashes, vague attributions, and chatbot leakage as the three tells worth fixing even once.
• No pattern numbers in bundle prose. SKILL.md, shared/persona/humanizer-steps.md, CLAUDE.md, and the slop-check hook's comments no longer cite patterns by number, since upstream renumbers them between releases.

Removed

• humanizer/references/patterns.md. Superseded by the vendored upstream catalog. Nothing is lost: every pattern it covered is in the new file, alongside seven more.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.13.1

What changed

Changed

• The GitHub repo is now betahope/cofounder-team. The rename completes the naming cleanup started in 0.13.0: repo, folder, README title, and skill names all say cofounder-team now. Every repo URL in README.md and claude-ai-project-instructions-template.md points at the new address. Old founding-team links keep working through GitHub's redirect, and existing clones need no action (the redirect covers git operations too).

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.13.0

What changed

The "10x" release: two new capabilities (a team skill and shared company memory), a smarter humanizer, and a round of content, tooling, and consistency upgrades across the bundle.

Added

• team skill. Convenes the co-founders in a single reply for cross-domain questions and big decisions: the most relevant co-founder leads, only the needed voices join, each is attributed by name, and every session ends in one clear recommendation. Portable, so it ships to Claude.ai, ChatGPT, and every other Agent Skills tool, bringing the multi-voice behavior that previously existed only in the Claude.ai project template to every platform. Added to the README skills list and install steps.
• Shared company memory (Claude Code only). The four personas, both coaches, and the team skill now share one memory file, ./.cofounder-team/company.md in the founder's project, holding the durable basics (stage, ICP, metrics, runway, burn, cap table, top goal). Skills read it at session start instead of re-asking, update it when facts change, and offer once to create it. Defined in shared/persona/company-memory.md inside a {{FLAVOR:claude-code}} block; the portable build never sees it. Documented as cross-skill contract #4 in CLAUDE.md.
• Humanizer: anti-fabrication rule. The workflow now forbids introducing facts, numbers, names, or citations not present in the source (placeholder or ask instead), and references/patterns.md opens with a note that the invented specifics in its "after" examples are illustrative, not licence to fabricate.
• Humanizer: density judgment and newer tells. New "judge density, not presence" rule (count occurrences against length; one em dash is typography, three per paragraph is a tell), a warning that the voice examples are one voice, not the voice (so humanized pieces stop converging on the same casual-blogger sound), and extended watch lists: verb slop (leverage, unlock, streamline, robust, game-changer, holistic, supercharge, elevate, empower, seamless), "here's the thing", and perfectly parallel identical-opener bullet lists.
• Missing weak/strong examples in startup-application-coach. The competitors, why-now, customer acquisition, and milestones questions in references/questions.md now each have a weak/strong pair (fictional companies, labelled as such). references/moats.md gains a worked false-moat-to-honest-sequencing example plus two missing moat types: counter-positioning and regulatory/compliance position.
• Missing slides in pitch-deck-coach. references/slide-by-slide.md gains a go-to-market slide section (one concrete channel, evidence it works, why it is hard to copy) and use-of-funds guidance on the ask slide (money-to-milestones lines, 18-to-24-month operating plan, and how to handle forms that demand 5-year projections).
• Freshness discipline on program facts. references/program-specific.md opens with a standing rule (figures dated mid-2026; trust the program's live application over this file) and notes YC's four-batches-a-year cadence. references/video.md hedges the verbatim YC instructions the same way and adds an explicit rule: real founders on camera, no AI avatars or synthetic voiceover, for both video types.

Changed

• Humanizer works on non-English drafts now, structurally. The old contract skipped the humanizer entirely for non-English copy. The new contract: lexical patterns (word lists, English idiom) remain English-only, but structural patterns (forced triplets, stacked hedges, signposting, label-colon-restatement lists, generic endings, chatbot leakage) apply in any language, so non-English drafts get a structural-only pass with a one-line note. Updated in the humanizer's scope section, the shared persona snippet, both coaches, the README, and CLAUDE.md contract #3.
• The build renders every .md file, not just SKILL.md. Reference files now go through the same directive pipeline (includes, flavor blocks, variables), enabling shared snippets inside references/. Files without directives render byte-identical to a plain copy. Documented in CLAUDE.md.
• Duplicated content moved to shared/ includes. The team-bio weak/strong example and the bottom-up TAM strong example (the two examples that had already drifted apart between files), both coaches' conversation-style and humanizer-language sections, and the personas' "no em dashes" / "default to brief" bullets now live in single shared snippets. The vision guidance intentionally stays two-level (summary in SKILL.md, detail in vision.md) rather than deduplicated, since the two serve different loading stages.
• The slop-check hook is quieter and sharper. It now scans only the text Claude just wrote (pre-existing text no longer re-triggers it on every edit), requires two kinds of tell (or a pile-up of AI words) before speaking, drops the lone en-dash and curly-apostrophe triggers, matches "let's dive" with either apostrophe, and syncs its word list with the humanizer's pattern catalog. hooks/README.md updated to match.
• Persona descriptions trimmed to what-and-when. The behavior summary sentence ("challenges weak assumptions... takes a clear position") is gone from all four personas' frontmatter descriptions; the domain keywords and triggers stay. The behavior itself is unchanged; it lives in each skill's "How you think" section, which the model now has no shortcut around.
• Bundle renamed to "The Co-Founder Team" in the README. Title and tagline now match the folder, skill names, and changelog. The GitHub repo URLs still point at betahope/founding-team until the repo itself is renamed (GitHub will redirect once it is).
• Application coach: replaced the unexplained "$2.5M in revenue" acquisition target in references/questions.md with stage-appropriate phrasing (first 100 users, first 20 paying customers, first $100K).

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.12.0

What changed

A correctness and consistency release: fixes from a full review of every skill against its own rules and the Agent Skills spec.

Fixed

• Bottom-up TAM math error. The worked example in startup-application-coach said 749,404 restaurants at $1,200/year is "roughly $749.4M". The correct product is roughly $899M. Fixed in references/bottom-up-tam.md and references/questions.md; pitch-deck-coach's copy of the example already had the right figure.
• Fabricated facts about real companies in example bios. The strong team-slide and team-answer examples credited "Jordan" with roles at Stripe and Brex, including a Brex acquisition in 2018 that never happened. Both examples now use clearly fictional companies, with a note that real answers should name real, verifiable employers. Fixed in pitch-deck-coach/references/slide-by-slide.md and startup-application-coach/references/questions.md.
• Broken repo links in the Claude.ai template. claude-ai-project-instructions-template.md linked to betahope/cofounder-team; the repo is betahope/founding-team. Both links 404'd.
• Logo self-contradiction in the design reference. design-principles.md told founders to put the logo on every slide in one section and called that an amateur mistake in another. Resolved to the cover-only rule, matching the skill's red-flag list.
• Short live pitches now carry the vision beat. The skill makes vision non-negotiable at every length, but the 1-minute, 3-minute, 5-minute, and 10-minute structures in deck-types.md never mentioned it. Each length now says where vision goes.
• Demo day length corrected. audience-asks.md and deck-types.md stated demo days are "usually 3 minutes". Formats vary by program (YC about 1 minute, Techstars about 5); both files now say so and tell founders to confirm with the program.
• Wrong relative links between reference files. Files inside references/ linked to siblings as references/<file>.md, which resolves nowhere from there. Now plain sibling filenames, per the Agent Skills spec's file-reference guidance. Fixed in pitch-deck-coach/references/slide-by-slide.md, startup-application-coach/references/questions.md, and references/program-specific.md.
• Grammar in Jack, Maya, and Priya's descriptions. "keep things as simple" is now "keeps things as simple" in all three frontmatter descriptions, the text every user sees in their skills list.
• The humanizer no longer uses em dashes in its own prose. The skill that flags em-dash overuse (and the repo that bans em dashes) was full of them, including one in a recommended "After" rewrite. Instructional prose in humanizer/SKILL.md, references/patterns.md, references/voice-calibration.md, and references/example.md is now em-dash-free; quoted example text that demonstrates AI tells keeps them on purpose.

Changed

• Dan's copy delivery is now flavor-aware. The rule to save final copy to ./drafts/<slug>.md is Claude Code only; the portable build (Claude.ai, ChatGPT, and other Agent Skills tools) now instead asks for final copy in a clearly separated Markdown block, with no dependence on file paths. Previously the portable build shipped a filesystem instruction that cannot work there.
• Persona section order unified. Maya, Priya, and Dan now place "How you talk" before the humanizer section, matching Jack and the canonical shape documented in CLAUDE.md. No wording changed; sections moved only.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.11.2

What changed

Added

• .gstack/ in .gitignore. Local gstack developer tooling creates this folder inside the repo; it is machine-local state and never part of the bundle, so git now ignores it.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.11.1

What changed

Makes the multi-tool story concrete: a hands-on tested ChatGPT install path in the README, and the internal build flavor renamed to match what it really is.

Added

• "Install in ChatGPT" section in README.md. ChatGPT supports the Agent Skills standard, and we verified our zips there by hand: open Plugins in the sidebar, switch to the Skills tab (or go to chatgpt.com/skills), click +, choose Upload from your computer. Upgrading is re-download and re-upload, the same as on Claude.ai.
• "Install in other tools" section in README.md. Covers Cursor, OpenAI Codex, GitHub Copilot, Gemini CLI, and other Agent Skills tools, with honest expectations: we test on Claude and ChatGPT, elsewhere the skills are expected to work rather than guaranteed.

Changed

• Renamed the claude-ai build flavor to portable. The words-only build now lives at dist/portable/ and the directive is {{FLAVOR:portable}}, since the release zips ship to Claude.ai, ChatGPT, and any other Agent Skills tool, not just the Claude website. Internal only: zip names, install steps, and installed skills are unchanged. build, the release workflow, and CLAUDE.md were updated together; no source skill used the old directive.
• README reframed as multi-tool. New tagline, install intro, and a slimmer "Built on an open standard" section that points at the install paths instead of repeating them.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).

v0.11.0

What changed

Aligns the skill bundle with the Agent Skills spec (https://agentskills.io/specification), the open standard shared by Claude Code, Claude.ai, and other tools.

Added

• Spec-compliant frontmatter. Every skill's frontmatter now carries license: MIT and a metadata block (author: betahope, bundle-version). The build reads the version from the VERSION file through a new built-in {{var:BUNDLE_VERSION}} directive, so a release zip identifies which bundle version it came from.
• compatibility field on cofounder-team-upgrade. Marks it Claude Code only, since it needs git and a local clone. This is the spec's intended use of that field.
• CI validation gate. The release workflow now validates every built skill, in both flavors, with the official skills-ref validator right after building and before zipping, so an off-spec skill can never ship.
• "Built on an open standard" section in README.md. The intro now calls the bundle portable Agent Skills, and the new section explains that release zips are plain Agent Skills folders that work in any tool supporting the standard (Cursor, OpenAI Codex, GitHub Copilot, Gemini CLI, and others), not just Claude Code and Claude.ai.
• Docs. CLAUDE.md documents the built-in {{var:BUNDLE_VERSION}} directive, the spec's allowed frontmatter fields, and where the humanizer version now lives.

Changed

• humanizer frontmatter, made spec-compliant. The top-level version: 3.0.0 field, not allowed by the spec and the one thing that failed the official validator, moved to metadata.version. allowed-tools changed from a YAML list to the spec's space-separated string. The compatibility field was dropped, since the skill is words-only and runs anywhere. The description was rewritten without em dashes.
• Dan's description trimmed. Down from exactly 1024 characters, the spec's hard limit, to 981, keeping every trigger keyword.

Install in Claude.ai

Download the zips for the skills you want. Then in Claude.ai, open Customize → Skills from the left sidebar, click Upload, and drop each zip in.

The cofounder-team-upgrade skill is Claude Code only and is not included here. To get a newer version on Claude.ai, download the latest release zips and re-upload.

For Claude Code, install the whole bundle in one command (see the README).