AI Tinkerers / local meetup field guide

Agent
Automations

Four practical demos that started as casual requests and became operating loops: capture, research, write, verify, publish, reconcile. Some began in OpenClaw. Some are already ported into the current Hermes/site workflow.

Demo thesis

The impressive part is not the prompt. It is the definition of done.

These automations are useful because they keep working past the generated text: they inspect the repo, touch real files, process assets, run builds, post API payloads, and leave behind verifiable artifacts.

5

demoable automations

4

content pipelines

3

production publishing loops

18

logged betting review days

01 human request
02 agent workflow
03 tool calls
04 verified artifact

The demo stack

Four ways agents stop being chatbots.

01

Ported to site workflow

No Reserv-ai-tions

A full-stack publishing flow for restaurant and venue reviews: casual request plus a photo in, structured review, processed assets, verified build, and production publish out.

Why demo it: Shows research, writing, asset processing, and deployment in one loop.

Input

  • • Natural-language request
  • • Venue name / context
  • • Real photo from the experience

Core tool calls

  • • read repo schema + nearby reviews
  • • web_search / web_fetch venue metadata
  • • write review markdown
  • • exec image processing + build + git

Workflow

  1. 01 Start in the Astro site repo
  2. 02 Check whether the venue already exists
  3. 03 Research address, coordinates, tags, and URL
  4. 04 Write review markdown in the site schema and voice
  5. 05 Process the image into site assets
  6. 06 Run the build, commit, rebase, and push

Source of truth

  • src/content/reviews/*.md
  • public/no-reserv-ai-tions/

Definition of done

  • ✓ Review file exists
  • ✓ Hero and thumb generated
  • ✓ Build passes
  • ✓ Changes pushed
02

Ported to site workflow

Music reviews

A metadata-heavy review generator that starts with a Spotify link or song request and ends with schema-valid editorial content.

Why demo it: Demonstrates structured metadata plus subjective writing in the same task.

Input

  • • Spotify track URL or song request
  • • Optional notes about why the track matters

Core tool calls

  • • read schema + existing song entries
  • • exec generation script + build
  • • write / edit final markdown

Workflow

  1. 01 Inspect existing song entries
  2. 02 Verify the songs collection schema
  3. 03 Use the Spotify URL as the source input
  4. 04 Run the generation script or follow its metadata model
  5. 05 Create or update src/content/songs markdown
  6. 06 Verify metadata completeness and run the build

Source of truth

  • src/content/songs/*.md
  • src/content.config.ts
  • scripts/generate-song-review.mjs

Definition of done

  • ✓ Song markdown exists
  • ✓ Frontmatter matches schema
  • ✓ Review body is complete
  • ✓ Build passes
03

Started in OpenClaw

Stash link capture

A low-friction Discord capture flow: drop a link in the stash channel, extract the URL and notes, POST to the stash API, and reply with a minimal receipt.

Why demo it: Shows invisible, useful automation embedded where the capture already happens.

Input

  • • Normal Discord channel message
  • • At least one URL
  • • Optional trailing notes

Core tool calls

  • • channel-specific prompt routing
  • • read local stash API key
  • • HTTP POST to stash API

Workflow

  1. 01 Message arrives in the stash channel
  2. 02 Agent checks for a URL
  3. 03 First URL is extracted
  4. 04 Remaining text becomes notes or null
  5. 05 Agent reads the API key and posts JSON
  6. 06 Agent replies with saved or failed status

Source of truth

  • OpenClaw config
  • Discord channel routing
  • https://stash-ecru.vercel.app/api/links

Definition of done

  • ✓ URL saved
  • ✓ Notes preserved
  • ✓ Success/failure response posted
04

Ported and evolving

Betting review + recap pipeline

A daily lifecycle that starts from screenshot-based pick capture, creates an Obsidian board, ranks recommendations, reconciles outcomes, and can publish a Plus EV recap.

Why demo it: Combines vision, verification, risk judgment, memory, and publishing over multiple turns.

Input

  • • Screenshots or typed betting picks
  • • Optional placed-bet notes
  • • Follow-up requests for review, settlement, or recap

Core tool calls

  • • image parsing for screenshots
  • • write / edit daily Obsidian note
  • • web verification of slate context
  • • exec for publishing checks + git

Workflow

  1. 01 Create or update the daily betting note
  2. 02 Parse screenshots into structured picks
  3. 03 Verify the actual slate and dates
  4. 04 Rank the board into Tier 1 / Tier 2 / Tier 3
  5. 05 Document placed bets separately
  6. 06 Reconcile outcomes and publish recap when useful

Source of truth

  • Obsidian/Claw/Betting/Daily/
  • Prompts/Daily Bet Review.md
  • src/content/plus-ev/

Definition of done

  • ✓ Daily note exists
  • ✓ Picks are cleaned
  • ✓ Recommendation uses the standard tiers
  • ✓ Outcomes or recap added when requested
05

Ported to Hermes workflow

TIL → YIL publishing flow

A two-step learning-to-page system: capture raw notes during the day, synthesize a design artifact the next morning, then build a reviewed Astro page from that spec.

Why demo it: Shows agents as editorial collaborators: they preserve messy learning, make design decisions explicit, then implement only after review.

Input

  • • Obsidian TIL inbox for a date
  • • Optional design-analysis note
  • • Approval to move from draft design to build

Core tool calls

  • • read Obsidian TIL inbox
  • • write DESIGN.md-style artifact
  • • inspect recent YIL page conventions
  • • astro check + build + git push

Workflow

  1. 01 Capture raw learnings in Obsidian throughout the day
  2. 02 Run the design step to produce a draft artifact under src/pages/yil/_design
  3. 03 Review the page shape, theme, Tailwind tokens, and open questions
  4. 04 Build src/pages/yil/YYYY-MM-DD.astro from the approved design
  5. 05 Export yilMeta so the YIL index discovers it automatically
  6. 06 Update series navigation, run Astro checks, commit, and push

Source of truth

  • Obsidian/Claw/TIL/Inbox/YYYY-MM-DD.md
  • src/pages/yil/_design/YYYY-MM-DD.md
  • src/pages/yil/YYYY-MM-DD.astro

Definition of done

  • ✓ Design artifact is draft until reviewed
  • ✓ Built page follows the artifact
  • ✓ yilMeta is exported
  • ✓ Astro check passes and changes are pushed

Betting harness receipts

The demo can answer the obvious follow-up.

Since May 1, the betting workflow has not just produced daily recommendations. It has kept a reconciled ledger in Obsidian, separating the Rithmm board from each agent harness and counting skips separately from picks.

Skip sweep footnote

May 7: 4/4 harnesses skipped

Rithmm board went 3-0

The cleanest “discipline had opportunity cost” example: the agents unanimously protected the bankroll while the board quietly swept.

Source

Win

Loss

Void / Skip

Rithmm visible board

Every reconciled visible candidate from the May 1+ Obsidian notes.

50
25
1 Void

ChatGPT harness

Strictest pass-enforcer profile; rarely converted a board into a pick.

3
1
14 Skip

Claude harness

Most action-oriented harness in the sample and strongest raw pick volume.

10
5
3 Skip

Gemini harness

Good pick record with one missing/not-logged day excluded from the count.

7
2
8 Skip

Perplexity harness

Useful context engine, but the forced-pick ledger has been rough so far.

3
6
7 Skip

Scoring rule

Harnesses only receive a win or loss when they made a pick. A pure pass, skip, or paper-only forced lean stays in the Skip column, which keeps discipline visible instead of burying it in the record.

Closing frame

Demo the loop, not the magic trick.

Start messy

screenshots, Discord messages, Spotify links, casual asks

Make it structured

schemas, notes, markdown, JSON payloads, ranked boards

Verify the result

builds, readbacks, commits, API receipts, settled outcomes