# The Agent Stack Playbook ### by FounderVibe Advisory **Stop collecting AI tools. Start running a stack.** --- > **A note on honesty, up front.** Every tool referenced in this playbook is open-source and free. You are not paying for the tools — you're paying for two years of evaluation compressed into two hours of reading: which components are worth your time, how they combine into working systems, and the operating discipline that makes them stick. Every recommendation links to its source with full attribution. > **[PUBLISHER TODO before selling]:** For every repo you list in Chapter 4, verify: (1) a LICENSE file exists and permits what you're describing, (2) the repo was updated in the last 6 months, (3) you have personally run it. Record name · link · license · date checked. Never list a repo you haven't run. --- ## Chapter 1 — The Stack Mindset The difference between founders who get leverage from AI and founders who get a bookmark folder is not tool knowledge. It's architecture. A **tool** answers a prompt. A **stack** runs a process. Tools require you to show up with context, instructions, and quality control every single time — which is why most founders' AI usage quietly collapses back to "write me a LinkedIn post." A stack holds the context permanently, encodes the instructions as workflows, and makes quality control a checklist instead of a vibe. The test: *if you disappeared for two weeks, would your AI setup produce anything?* If no, you have tools. This playbook is about crossing that line. **The three failure modes this book prevents:** 1. **The Collector** — 40 tools bookmarked, 3 used, 0 systematized. Cure: the curation framework (Ch. 3). 2. **The Prompt Hamster** — everything works, nothing is saved; every task starts from zero. Cure: the knowledge layer (Ch. 2, Layer 4). 3. **The Broken Robot** — an ambitious agent setup that failed once, was never debugged, and now everything is manual again. Cure: runbooks and the adoption plan (Ch. 6–7). ## Chapter 2 — Anatomy of an Agent Stack Every working stack has four layers. Miss one and the system decays. ``` ┌──────────────────────────────────────────────┐ │ L4 · KNOWLEDGE context files, prompt │ │ (memory) library, SOPs — versioned │ ├──────────────────────────────────────────────┤ │ L3 · ORCHESTRATION workflows, schedules, │ │ (process) handoffs between agents │ ├──────────────────────────────────────────────┤ │ L2 · AGENTS & SKILLS the specialists: │ │ (capability) research, content, │ │ analytics, ops │ ├──────────────────────────────────────────────┤ │ L1 · FOUNDATION the model + runtime │ │ (engine) (your AI provider account) │ └──────────────────────────────────────────────┘ ``` **L1 — Foundation.** Your model access (API account with a major provider) and a runtime for agents (a terminal-based agent framework or an agent-capable desktop app). Cost reality: €20–100/month usage-based. Own this account yourself; never run your business through someone else's key. **L2 — Agents & Skills.** The open-source ecosystem's gift: pre-built specialists. A "skill" is a packaged capability (e.g., competitor research done a specific way); an "agent" is a persona plus skills plus autonomy. This layer is where curation matters most — and where most repos are abandoned experiments. Chapter 3 is your filter. **L3 — Orchestration.** The difference between "I have agents" and "work happens." A workflow = trigger + sequence + output + owner. Written down. Anything not written down doesn't exist on a busy Tuesday. **L4 — Knowledge.** The layer everyone skips and the reason everything else compounds. Four files, versioned, loaded into every agent session: - `positioning.md` — what you are, for whom, against what - `icp.md` — who buys, their pain in their words - `voice.md` — how you sound; words you never use - `offers.md` — your ladder and prices Change your positioning once, and every downstream output changes with it. This is the layer that makes the stack *yours*. ## Chapter 3 — The Curation Framework GitHub has thousands of agent repos. Perhaps two percent deserve a place in a business stack. Run every candidate through five gates, in order — most die at gate one, which is the point. **Gate 1 · License.** Open the LICENSE file first, not the README. No license = all rights reserved = not usable in your business, full stop. MIT/Apache-2.0: use freely, keep attribution. GPL-family: fine to *use*, know the obligations if you redistribute. NC-clauses: not for commercial use — skip. **Gate 2 · Pulse.** Last commit within 6 months? Issues getting answered? A brilliant repo that's abandoned is technical debt with a nice README. **Gate 3 · Runability.** Can you get it working in under 30 minutes following only the README? If setup requires archaeology, operation will too. **Gate 4 · Fit.** Does it map to a workflow you actually run weekly? "Cool" is not a use case. Score against your bottleneck list, not your curiosity. **Gate 5 · Redundancy.** Prefer the boring component that overlaps your stack's patterns over the exciting one that introduces a new runtime. Every additional pattern is a future 2 a.m. debugging session. **The Stack Ledger.** Keep one file: component · repo link · license · version · date checked · which workflow uses it. Ten minutes of bookkeeping that saves you from both legal surprises and mystery breakage. ## Chapter 4 — The Marketing Stack Blueprint The reference build. Five roles, filled from curated open-source components. > **[PUBLISHER TODO]:** insert your verified picks per role, each as: **Name** · repo link · license · one honest sentence on why it earned the slot and one on its weakness. 1–3 picks per role, no more. **Role 1 · The CMO.** A strategy-level agent loaded with all four L4 context files. Job: campaign planning, prioritization, reviewing the specialists' output against positioning. This is the role where persona-based agent packs shine. → Picks: `[INSERT: verified CMO/persona agent repo(s)]` **Role 2 · The Researcher.** Market scans, competitor monitoring, voice-of-customer mining. Output standard: findings with sources, never vibes. → Picks: `[INSERT: verified research skill repo(s)]` **Role 3 · The Content Engine.** Long-form drafts, repurposing pipelines (one pillar piece → posts, email, threads), always writing *from* `voice.md`, never from model defaults. → Picks: `[INSERT: verified content/marketing-skills repo(s)]` **Role 4 · The Analyst.** A dashboard answering three questions weekly: what did we ship, what did it do, what dies next week. → Picks: `[INSERT: verified dashboard repo(s)]` **Role 5 · The Librarian (you, 20 min/week).** Updates context files, saves winning prompts to the library, maintains the Stack Ledger. Unautomatable and non-negotiable — this role is why the stack still works in month six. ## Chapter 5 — Three Workflows That Pay For This Book Each follows the same spec format: **Trigger → Steps → Output → Owner → Time.** **Workflow A · The Weekly Content Sprint (90 min → a week of content)** Trigger: Monday, 9:00. Steps: (1) CMO agent proposes 3 angles from `positioning.md` + this week's goals; you pick one in 5 minutes. (2) Researcher pulls supporting evidence and customer language. (3) Content Engine drafts the pillar piece; you edit — editing is the human job, first drafts are not. (4) Repurposing pass: platform-native variants, not copy-paste. Output: 1 pillar + 4–6 derivatives, scheduled. Owner: operator. **Workflow B · The Competitor Pulse (30 min/month, runs itself)** Trigger: 1st of month. Steps: Researcher sweeps competitor sites, changelogs, reviews → diffs against last month → CMO agent annotates: what threatens us, what validates us, what do we ignore. Output: one-page brief in your inbox. Owner: CMO agent; you read it with coffee. **Workflow C · The Launch Assembly Line (per launch, 1 day instead of 1 week)** Trigger: anything new ships. Steps: feed the offer spec → CMO drafts the launch plan → Content Engine produces page copy, emails, and posts from `offers.md` + `voice.md` → you review once, as a batch. Output: complete launch kit. Owner: you, because launches deserve a human at the wheel. ## Chapter 6 — Operations: Keeping It Alive - **Runbooks or it didn't happen.** One page per workflow: purpose, trigger, steps, expected output, failure modes. Written for a smart stranger. - **The Friday Five.** Five questions, five minutes, every week: What ran? What broke? What output was embarrassing? What prompt deserves the library? What component hasn't earned its keep this month? - **Version your knowledge layer.** Date-stamp context file changes. When output quality shifts, you'll know why. - **Prune quarterly.** A stack is defined by what you remove. Unused components are attack surface and confusion; cut without sentiment. - **Never automate judgment.** Pricing, strategy, apologies, anything legal — human-only, forever. Write this list down before you need it. ## Chapter 7 — The 30-Day Adoption Plan - **Week 1 · Foundation.** L1 running, all four context files written (worst draft beats no draft), Stack Ledger started. No agents yet — context first is the whole trick. - **Week 2 · First hire.** Install ONE agent for your biggest bottleneck. Run Workflow A end-to-end twice, supervised. Write its runbook. - **Week 3 · Second hire.** Add the Researcher + Workflow B. Start the prompt library with your five best. - **Week 4 · Review like a manager.** Run the Friday Five on the whole month. Cut anything unused. Decide the next single addition — one per month from here, maximum. The failure mode to avoid is installing everything in week one and abandoning everything in week three. One component at a time, each earning its place — that's how a stack becomes infrastructure instead of a phase. --- ## Appendix A — The Stack Ledger template ``` COMPONENT | REPO/LINK | LICENSE | VERSION | CHECKED | USED BY WORKFLOW ----------|-----------|---------|---------|---------|------------------ | | | | | ``` ## Appendix B — Context file skeletons **positioning.md:** We are [category] for [ICP]. Unlike [alternative], we [difference]. We win when the buyer cares about [value]. **icp.md:** Role/situation · trigger moment · pain in their words (3 quotes) · objections · where they gather. **voice.md:** 3 adjectives · sentence rhythm · always-words · never-words · 2 sample paragraphs that sound like us. **offers.md:** ladder table: product · promise · price · next step up. ## Appendix C — Never-automate list (starter) Pricing decisions · strategy pivots · legal/compliance answers · apologies and complaints · hiring/firing communication · anything you'd want to have written yourself in a crisis. --- *The Agent Stack Playbook v1.0 · FounderVibe Advisory · All referenced tools are open source and attributed; verify licenses before commercial use in your own context.*