Every founder I know has asked an LLM to "review my pitch deck". The answer is always the same: polite, generic, and useless. "Consider adding more detail to your traction slide." Thanks.
The model is not the problem. The problem is that the model has no idea what a good deck looks like to an investor, and nobody tells it. So I spent a few weeks reading what the people who actually fund companies have published on the topic, distilled it into an Agent Skill, and open-sourced it.
npx skills add amaklakov-droid/pitch-deck-skill
That one line installs it into Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, or whichever agents you pick. The repository is amaklakov-droid/pitch-deck-skill, MIT licensed.
This post covers what the skill knows, why I built it as a skill rather than a prompt, and what a review session looks like.
The one number that changes how you write a deck
DocSend has tracked how investors read decks across more than 200,000 interactions. The median first read takes 3 minutes 44 seconds. Decks that fail to raise lose the reader at around 2 minutes 13 seconds. Top funds fund roughly 1 of every 200 warm pitches.
Once you internalize that, most deck advice collapses into one rule: a deck is not built to explain everything. It is built to be understood and remembered in three minutes and to earn a meeting. The meeting, not the investment, is the deck's only job.
That single principle is the spine of the skill. Every rule in it traces back to the 3-minute read.
What went into it
The skill distills four sources that mostly agree with each other, which is itself a useful signal:
- Y Combinator, Kevin Hale's "Legible, Simple, Obvious" design laws and the YC library on pitching.
- Sequoia Capital's business-plan template, the outline behind the famous Airbnb seed deck.
- DocSend research on reading time, slide order, and which slides investors actually stop on (Team and Traction, by a wide margin).
- Deep-tech funds: Khosla Ventures, Julian Capital's Deep Checks, Hard2beat. Hardware, biotech, materials and energy decks answer a different risk question ("will the physics work and scale?") and need extra slides most templates skip.
I wrote the research up first as a long analytical note, then compressed it into the skill. The note is in the repo if you want the sources with context.
What the skill actually does
Three modes, picked automatically from what you ask.
Create. You describe the company, stage, and audience. The skill picks the 5 to 7 ideas an investor must remember, states them back to you, then produces a slide-by-slide outline. Every slide gets an assertion headline ("Contractors lose $400 a week to scheduling chaos", not "Problem"), body content under about 50 words, and a speaker note. Stage matters: at pre-seed the Why Now slide is critical, at Series A it is barely read, so the outline shifts accordingly.
Review. You attach a deck. The skill runs the 3-minute test first: after a short read, can a stranger retell what you do, for whom, why now, what proves it, and how much you are raising? Then it goes slide by slide and sorts findings into critical red flags, structural gaps, and form issues. Every finding comes with a rewrite, not just a complaint.
Slide surgery. "Rewrite my problem slide, investors keep zoning out." One slide, in and out.
The red-flag list
This is the part founders find most useful, because these are the things nobody tells you until an investor has already passed. The skill refuses to add any of them when creating and flags them as critical when reviewing. A sample:
| Red flag | Why it kills the deck |
|---|---|
| NDA request before viewing | Marks you as a novice. Investors do not sign pitch-stage NDAs. |
| Exit-strategy slide | Great companies are bought, not sold. Its absence is one of the most stable markers of strong decks. |
| "We have no competitors" | Reads as an unresearched market. Excel, manual work and "do nothing" are competitors. |
| Valuation in the deck | Valuation is negotiated, never presented. |
| Ask as a range | One number plus a specific use of funds. |
| "18 months of runway" as the plan | Runway is not a plan. Say what becomes true about the company. |
| Vanity metrics | Signups and page views instead of revenue, retention, cohorts. |
| Top-down market sizing | "1% of $500B" with no bottom-up arithmetic. |
The full list has 18 entries, including form violations like the 50-word slide budget and the "send a PDF, never a Drive link with an access request" rule.
Why a skill and not a prompt
I tried the prompt route first. It fails for three reasons.
Context cost. The full guidance is thousands of words. Pasting it into every conversation is expensive and most of it is irrelevant to any given question. Skills use progressive disclosure: the agent loads a compact SKILL.md when the topic comes up and reads the reference files (per-slide guide, deep-tech specifics, review checklist) only when needed.
Triggering. A prompt has to be remembered. A skill has a description that the agent matches against the conversation, so it activates when you say "help me raise a seed round" or "prepare for demo day" without you invoking anything. I included trigger phrases in Russian too, since that is where half my users are.
Portability. Anthropic published the SKILL.md format as an open standard in December 2025 and by now roughly forty products support it. One folder of Markdown works across Claude Code, Codex, Cursor, Copilot, Gemini CLI and the rest. A prompt lives in one chat window.
The layout is deliberately boring:
pitch-deck-advisor/
├── SKILL.md # workflow, principles, 12-slide canon, red flags
└── references/
├── slide-guide.md # per-slide patterns and failure modes
├── deep-tech.md # hardware / bio / materials / energy specifics
└── review-checklist.md # 18 red flags, stage calibration, known-deck lessons
Plain Markdown, no scripts, no dependencies. You can read the whole thing in ten minutes and fork it if you disagree with a rule.
A session, roughly
You: "Review this deck and be brutal." (deck attached, battery-materials startup, seed)
The skill asks nothing if the deck answers the basics, otherwise one short question about stage and audience. Then:
3-minute test: fails 2 of 5. A stranger can retell what you do and for whom. They cannot retell why now (no slide) or what proves it (slide 7 shows a pilot logo wall with no numbers).
Critical: slide 11 states a $40M pre-money valuation. Remove it. Valuation is negotiated.
Structural: no Why Now. At seed this is important. Proposed headline: "Solid-state cell makers hit the cathode cost wall in 2025; three of the top five have public 2027 targets they cannot meet with current materials."
Deep tech: the technology slide mixes proven and to-prove claims. Split them: what is validated at what scale, what the round de-risks, with the milestone that proves it.
And so on through the deck, every item with replacement text. The output is opinionated on purpose. If you want a cheerleader, the default model is already one.
Install
Any agent supporting the Agent Skills standard:
npx skills add amaklakov-droid/pitch-deck-skill
Claude Code, as a plugin:
/plugin marketplace add amaklakov-droid/pitch-deck-skill
/plugin install pitch-deck-advisor@pitch-deck-skill
Claude.ai or Claude Desktop: zip the pitch-deck-advisor folder and upload it under Settings, Capabilities, Skills.
Anything else: SKILL.md is a system prompt, the three reference files are attachments. It works, just without the automatic triggering.
What is next
The skill is at v1.0.0. Things I would like help with, in order:
- Eval cases. Real decks (anonymized) with a known investor outcome, so the review mode can be scored rather than vibe-checked.
- Fund-specific guidance. If you have pitched a fund with a distinctive format (a16z's memo culture, for instance), a reference file for it would be welcome.
- More languages for the README. The skill itself stays in English for cross-model reliability; it already answers in whatever language you write in.
Issues and PRs at amaklakov-droid/pitch-deck-skill. If it saves you one bad investor meeting, tell me which slide it fixed.
Top comments (0)