> ## Documentation Index
> Fetch the complete guide index at: https://www.synscribe.com/agentic-discovery/llms.txt
> Use this file to discover all pages before exploring further.

---
title: "The Agentic Discovery Playbook: 12 Plays, Sequenced"
description: "all 12 plays to become AI agents' default choice. What each does, who needs it, impact and effort, in one ten-minute overview with sequencing."
slug: /agentic-discovery/agentic-discovery-playbook
series: The Agentic Discovery Playbook · Part 4 of 6
last_verified: 2026-06-12
---

# The 12 Plays of Agentic Discovery: Overview and Sequencing

> **In short:** An agent runs four steps to pick a product: **find → research → shortlist → act**. Most plays are tools that serve more than one step, not items in a single box. **Find** (appear in its searches and training data): agent web search, registries, public evals. **Research** (get opened and read): llms.txt, markdown docs, snippet density. **Shortlist** (win the comparison): MCP server, skills & AGENTS.md, snippets, docs. **Act** (make integration effortless, and lock in for next time): directives, scaffolder rules, keyless onboarding. Find is the entry gate. The later stages move the final pick the most.

![The agentic-discovery territory: a matrix of who is choosing (coding agents like Claude Code, Codex, OpenCode; work agents; consumer agents like ChatGPT and Perplexity; and vertical enterprise “custobots”) against what you do to win each step of the find, research, shortlist, and act flow. The coding-agent column maps to the twelve plays.](/agentic-discovery/images/diagram-discovery-matrix.svg "The play map: what you do to win each stage, across every kind of agent.")

This page is the whole method in ten minutes. Each play links to a full-depth page with implementation steps, field examples, counterexamples, and a pass/fail eval. If you only skim one page in this guide, skim this one.

## Do this now

- [ ] Identify which of the four stages you're weakest on (most teams: all four)
- [ ] Run the [30-minute audit](/agentic-discovery/agent-readiness-audit) to get your baseline score
- [ ] Pick your track below (new product / established player / platform) and queue plays in that order
- [ ] Ship the week-one pair: registry entries (Play 2) + llms.txt (Play 5). Also make your category page easy to back up for agent search ([Play 1](/agentic-discovery/ai-agent-web-search-and-fetch))
- [ ] Put the [weekly tracker](/agentic-discovery/measure-ai-visibility) in place before you change anything. You'll want the before/after

## The four stages: live the agent's journey

An AI agent runs the same four steps every time it picks a product. The important thing: most of the twelve plays are **tools, not stages**. The same artifact (your docs, your registry entry, your evals) does a different job at each step. So plays recur across stages instead of living in one box.

### 1 · Find: get into the running

**The agent:** looks for candidates in what it was trained on (its prior) *and* in a live web fan-out: many parallel queries at once.

**You:** appear in its training data (the long game) and in those searches (now). Its fan-out queries are a free keyword map. Classic SEO/GEO still works here.

*Plays in play:* Agent web search · Registries & directories · Evals & leaderboards.

### 2 · Research: get opened and read

**The agent:** skims the result list and opens only a few, mostly off the search-result summary. The question it's answering: *"is this potentially what I want?"*

**You:** win the title and description, get backed up across third parties, be machine-readable, and cover enough use cases to match more of its queries.

*Plays in play:* Agent web search · Registries & directories · llms.txt · Markdown docs · Snippet density.

### 3 · Shortlist: win the comparison

**The agent:** now has a few real options and judges each on three things. *Does it do what it claims? Can I implement it easily? Is it the best of the lot?* This is the one moment it's comparing you head-to-head against the competitor docs sitting right beside you in context.

**You:** prove the capability, prove low-friction implementation, and prove you beat the alternatives.

*Plays in play:* Snippet density · MCP server · Markdown docs · Skills & AGENTS.md.

### 4 · Act: make integration effortless (and lock in)

**The agent:** has chosen. Now it integrates and uses you. You want an early win so it knows it's on track and doesn't second-guess back to a competitor.

**You:** make the first integration correct and frictionless, steer it past your own deprecated patterns, and persist yourself into the project so you start as the default next time.

*Plays in play:* Markdown docs · Directive layer · Skills & AGENTS.md · Scaffolder rules / CLAUDE.md · Keyless onboarding · Evals & leaderboards.

**The loop that compounds.** A few Act-stage plays (scaffolder rules, CLAUDE.md, AGENTS.md) don't just win *this* integration. They get written into the project and re-load next session as the agent's prior and environment. **Act feeds the next Find.** That feedback loop is why defaults compound. It's also why the environment plays are the strongest selection lever we measured.

### Which play serves which stage

Read the rows: most plays are doing work at more than one step.

| Play | Find | Research | Shortlist | Act |
|---|:--:|:--:|:--:|:--:|
| Agent web search | ● | ● | | |
| Registries & directories | ● | ● | | |
| MCP server | | | ● | |
| Skills & AGENTS.md | | | ● | ● |
| llms.txt | | ● | | |
| Markdown docs | | ● | ● | ● |
| Snippet density | | ● | ● | |
| Directive layer | | | | ● |
| Scaffolder rules / CLAUDE.md | | | | ● |
| Keyless onboarding | | | | ● |
| Evals & leaderboards | ● | | | ● |
| Agentic Resource Discovery (ARD) · *frontier* | ● | | | |

**Find is the entry gate.** You can't be researched, shortlisted, or integrated if the agent never finds you, and it's the widest, least-owned stage. But for *changing the final pick*, our pilots rank the later stages highest.

## The twelve plays

Each play once, with the stages it works in. Run order and tracks come after.

**[Play 1: Get found in the agent's own web search](/agentic-discovery/ai-agent-web-search-and-fetch).** *Works in: Find, Research.*

Before an agent reaches your structured docs it searches the open web itself. It opens only ~6% of what it surfaces, and kills ~48% of the self-reported claims it checks. Win the agent's own (long, dated, spec-loaded) query, be worth opening, and make every headline claim something a primary source can back up. *Who:* everyone; it's the entry gate. *Effort:* ongoing content discipline, no engineering. *Impact:* where the training prior gets overturned and newcomers first break in.

**[Play 2: Registries & directories](/agentic-discovery/ai-agent-registries-and-directories).** *Works in: Find, Research.*

Only four placements show measured impact (the Official MCP Registry chain into VS Code, skills.sh, Anthropic's surfaces, Context7). About 99% of real distribution bypasses the rest. They mostly help you get *indexed* (so you surface in search) and *backed up* (a second party naming you). Claim and curate the four, write your descriptions around task phrases, skip the junk tier. *Who:* everyone. *Effort:* 1–2 days, then 1 hr/week. *Expected impact:* immediate findability. The description-match win is free in uncontested categories.

**[Play 3: MCP server distribution](/agentic-discovery/mcp-server-distribution).** *Works in: Shortlist.*

A docs-search MCP first, then a live server with a sandbox endpoint. Its job is to prove you're easy to implement at the comparison step. An agent that can see you ship an MCP reads it as "this one is easy for me to use." The server is the easy half. Distribution (registry publish, `.well-known` manifests, per-client install snippets) decides whether it ever gets used. *Who:* API products. *Effort:* 1–3 weeks. *Impact:* in our pilots, agent-operability flipped *selection* when the agent knew about it.

**[Play 4: Agent skills & AGENTS.md](/agentic-discovery/agent-skills-and-agents-md).** *Works in: Shortlist, Act.*

A public repo of SKILL.md workflows makes you installable across ~55 agent clients with no gatekeeper. The proven starter set is best-practices + quickstart + migrate-from-competitor. Skills and AGENTS.md live *inside the agent's context*. At shortlist they signal "easy to implement." During Act they steer the integration itself. *Who:* devtools with conventions worth encoding. *Effort:* ~1 day for the starter set. *Impact:* vendors' best-practices skills run 100K+ installs. Your competitors may already be there.

**[Play 5: llms.txt](/agentic-discovery/llms-txt).** *Works in: Research.*

The index file agents and retrieval tools fetch first: sectioned links with one-line descriptions, tiered variants, and the part most guides miss, a directive section. It's what makes you clean to *read* the moment the agent opens you. Honest status: verified infrastructure, unproven ranking magic. *Who:* everyone. *Effort:* hours (free if your docs platform generates it).

**[Play 6: Markdown docs](/agentic-discovery/markdown-docs-for-ai-agents).** *Works in: Research, Shortlist, Act.*

Every docs URL serves raw markdown at `.md`. The gold standard also content-negotiates on canonical URLs. Frontmatter, agent banners, machine-readable changelog. Clean docs are read at Research, compared at Shortlist, and become the implementation reference at Act. It's the one play that runs through three stages. *Who:* everyone with docs. *Effort:* hours-to-days. Platform choice can make it free.

**[Play 7: Snippet engineering](/agentic-discovery/code-snippets-for-ai-agents).** *Works in: Research, Shortlist.*

Retrieval indexes score your docs on five published metrics. Density beats bulk: a 440-snippet library outscores a 2,297-snippet one by 18 points. Self-contained, task-shaped, deduplicated snippets get you read at Research and prove capability at Shortlist. *Who:* everyone. It's the biggest lever on retrieval quality scores. *Effort:* 2–4 weeks of docs work.

**[Play 8: The directive layer](/agentic-discovery/stop-ai-using-deprecated-apis).** *Works in: Act.*

Agents emit your deprecated APIs because models memorized them. ALWAYS/NEVER directives steer the agent at *write time*, past its stale prior *and* past any competing source in its context. They cover exactly the *stale window* between your API change and the next training run. In our pilot: 100% wrong → 0% wrong on a recent breaking change, and zero effect on ancient history. ~40 tokens per directive. *Who:* anyone who has shipped a breaking change in ~18 months. *Effort:* days, plus a CI eval.

**[Play 9: Scaffolder rules & CLAUDE.md](/agentic-discovery/scaffolder-rules-claude-md).** *Works in: Act (and the lock-in loop).*

The strongest lever we measured: rules files persist in the repo and bias every future agent session. Written during the act of setup, they pay off on the *next* cycle. They re-load as the prior and environment, so you start the next Find as the default. Bun's CLI writes them at `bun init`, disclosed and opt-out-able. Our trials: 100% choice flip. Run it ethically or it backfires. *Who:* products with a CLI/scaffolder touchpoint (soft variants exist for everyone else). *Effort:* days.

**[Play 10: Agent-first onboarding](/agentic-discovery/agent-first-onboarding).** *Works in: Act.*

Remove every human step between "agent discovers you" and "passing integration test": keyless modes, no-signup provisioning endpoints, sandbox-by-default keys. This is the early win. The agent integrates you end-to-end without stalling and second-guessing. Target: first successful API call in under five minutes, zero human actions. *Who:* API/SaaS products. *Effort:* product work, weeks.

**[Play 11: Evals & leaderboards](/agentic-discovery/ai-evals-and-leaderboards).** *Works in: Find, Act.*

Build the internal eval harness (deprecation, flip-rate, integration-completion, doc-QA) and gate docs changes on it. That's the Act-stage tuning loop. Then publish it as a reproducible public leaderboard: an independent, primary source that ranks in search and survives the research-stage verification cut (Find). QA tool, tuning loop, and citation magnet in one. *Who:* teams ready to compound. *Effort:* quarter-scale.

**[Play 12: Agentic Resource Discovery (ARD)](/agentic-discovery/agentic-resource-discovery).** *Works in: Find (frontier).*

Google's new `ai-catalog.json` spec: publish a `/.well-known/ai-catalog.json` describing your MCP servers, agents, and APIs so registries can index them and agents can find them by capability. This is the one play in the guide that is **not yet evidence-backed**: it's a three-week-old v0.9 draft, and our day-one census found 0 of 39 top sites (including all 11 companies that wrote the spec) publishing one. *Who:* anyone shipping a callable resource (MCP server, A2A agent, API). *Effort:* ~1 hour to ship the file. *Impact:* unproven; a cheap first-mover hedge, sequenced last on purpose. Run Plays 1–11 first.

## What order should you run the plays in?

| Track | Weeks 1–2 | Weeks 2–6 | Quarter 2+ |
|---|---|---|---|
| **New product breaking in** | **Play 1** + 2 + 5 (claim the empty space and be easy to back up in agent search: your entry gate) | 7 → 8 → 4 | 3 → 9 → 11 |
| **Established player with training-data baggage** | Play 8 first (your old APIs are the threat) + 2 | **1** → 5 → 6 → 7 (you're most exposed to the verification cut) | 3 → 10 → 11 |
| **Platform / framework** | Play 2 + 5 | **1** → 4 → 9 (scaffolder is your unfair advantage) | 10 → 11 |

## If you only do three things

1. **Claim your four Tier-1 registry entries and rewrite the descriptions around task phrases** ([Play 2](/agentic-discovery/ai-agent-registries-and-directories)). Highest leverage per hour.
2. **Write the directive section for everything you've deprecated since 2024** ([Play 8](/agentic-discovery/stop-ai-using-deprecated-apis)). It's the only play that fixes agents being *confidently wrong* about you.
3. **Ship a best-practices skill** ([Play 4](/agentic-discovery/agent-skills-and-agents-md)). One day of work, distribution across ~55 agent clients, and the pattern your competitors are already running.

## The receipts

The later stages move the final pick the most: **Act > Shortlist > Research > Find**. That ordering comes from our pilot experiments (single model, n=2–3 per arm, full designs and limitations in [Part 2](/agentic-discovery/how-ai-agents-choose-products) and the [Data Room](/agentic-discovery/data)). A rules file flipped product selection 3/3 vs 0/3 control. A 5-line directive cut deprecated-pattern emission from 2/2 to 0/2. An operability fact flipped a recommendation 2/2. Retrieval-layer stats (n=17 audited entries): freshness is the strongest quality-score correlate (ρ=−0.54); corpus mass barely matters (ρ≈0.24–0.30). Field validation: the products running the full stack (measured install counts, scaffolder defaults, eval-tuned rules) are documented in [Part 2](/agentic-discovery/how-ai-agents-choose-products)'s teardowns.

## FAQ

**What is the agentic discovery playbook?**
Eleven evidence-backed plays for making a product the default choice of AI agents, plus a twelfth frontier play (ARD) we're tracking but don't yet have data on. They're organized by the four steps an agent takes: **find** (appear in its searches and training data), **research** (get opened and read), **shortlist** (win the head-to-head comparison), and **act** (make integration effortless, and persist into its environment for next time). Most plays serve more than one step.

**Which play has the highest ROI?**
Per hour invested: registry claiming and description engineering (Play 2). Per absolute impact on agent selection in our pilots: rules files in the agent's environment (Play 9), which flipped choices 100% of the time.

**Do I need all twelve plays?**
No. Run the audit, pick your track, and ship the week-one pair first. Plays 9–11 assume product surface (a CLI, an API) that not every company has; the soft variants are listed in each play.

**How long until results?**
Registry and retrieval changes surface within one re-parse cycle (days). Environment-layer effects apply to every new project immediately. Measurement setup (Part 5) should precede everything so you can prove it.

---

*Last verified 2026-06-11. We re-test the claims on this page quarterly. Changes are logged in the [Data Room](/agentic-discovery/data).*

**Part of [The Complete Playbook to Agentic Discovery](/agentic-discovery).**

← Previous: [The Agent-Readiness Audit](/agentic-discovery/agent-readiness-audit) · Next: [Play 1: Get Found in the Agent's Web Search](/agentic-discovery/ai-agent-web-search-and-fetch) →

> **Stay ahead of the agents.** We re-test this playbook quarterly and publish what changed: new data, busted myths, ranking shifts. [Get the update digest →](/agentic-discovery#updates)
>
> **Want this done for you?** Synscribe runs agentic-discovery programs for B2B SaaS and developer platforms. [Talk to us →](/contact)