The End-to-End Workflow
Every trigger, node, skill and agent wired as one live graph, from a form fill to a delivered asset.
Nothing leaves the graph without a human yes.
Tools on this page
- Telegram
- n8n
- Claude Code
- VS Code
- Codex
- GitHub
- Vercel
A · TriggersWhere work enters.
- Message to Hermes (Telegram), then Read the intent
- Schedule (sweeps on a timer), then Read the intent
- Client event (n8n webhook), then Webhook in
- Read the intent (routes the work), then Project rules if build work; Router if a skill job
B · Knowledge and memoryChecked before anything is built.
- Project rules (CLAUDE.md), then My notes
- My notes (Hermes reads them), then Knowledge base
- Knowledge base (1,277 chunks), then Past work
- Past work (what shipped before), then Outcome
C · Frame the taskSpec before prompt.
- Outcome (what done looks like), then Constraints
- Constraints (limits and no-gos), then Files that matter
- Files that matter (context is a budget), then Acceptance criteria
- Acceptance criteria (how it gets graded), then Spec
- Spec (one brief per run), then Pick one small piece
D · Build loopOne small piece at a time.
- Pick one small piece (one task, one context), then Claude Code builds
- Claude Code builds (the builder), then Subagents in parallel
- Subagents in parallel (parts run at once), then VS Code takes over
- VS Code takes over (precision over speed), then Run it
E · Check and repeatNothing ships until it passes.
- Run it (tests and a real run), then Grade the result
- Grade the result (against the criteria), then Codex second opinion
- Codex second opinion (a fresh model reviews), then Pass?
- Pass? (three ways out), then Claude Code builds if fails; Bug or failure if breaks; Approval gate if passes
F · Debug and fixFind the cause. Never patch blind.
- Bug or failure (reproduce it first), then Rank the causes
- Rank the causes (likeliest first), then Test to confirm
- Test to confirm (prove the cause), then Fix and re-check
- Fix and re-check (back through Check), then Run it if re-check
G · Agents and skillsOne job each.
- Router (picks the specialist), then Researcher; Auditor; Builder; Writer; Reel maker; Reviewer; Deliverables; Monitor; Data tidy; Memory keeper
- Researcher (finds sources), then Self-check
- Auditor (n8n audits), then Self-check
- Builder (workflows), then Self-check
- Writer (articles), then Self-check
- Reel maker (scripts to renders), then Self-check
- Reviewer (second pair of eyes), then Self-check
- Deliverables (client-ready files), then Self-check
- Monitor (watches live systems), then Self-check
- Data tidy (clean records), then Self-check
- Memory keeper (saves what was learned), then Self-check
- Self-check (graded before the gate), then Approval gate
H · n8n automationPredictable, logged, reversible.
- Webhook in (n8n), then Wait
- Wait (until the right time), then Branch
- Branch (needs a yes?), then Approval gate if needs a yes; Action if routine
- Action (does the work), then Log
- Log (every run recorded), then Write back
I · Approval and deliveryA person says yes.
- Approval gate (approve or send back), then Deliver if approved; Claude Code builds if send back with a note
- Deliver (with the handover doc), then Commit if code; Write back
J · Git and deployShip, then prove it’s live.
- Commit (small, clear commits), then Push
- Push (to GitHub), then Deploy
- Deploy (Vercel and my VPS), then Verify live
- Verify live (on the real URL), then Save progress; Deploy if fails: roll back and redeploy
K · Write-back and resetThe next run starts smarter.
- Write back (to the contact record), then Save progress
- Save progress (watcher every 15 min), then Fold back
- Fold back (mistakes become rules), then Fresh session; Outcome if becomes a rule
- Fresh session (reset early and often)
Every term on the map, in one line.
Plain English for the technical words above.
- n8n
- The automation tool the client workflows run in: triggers, waits, branches and actions, wired as nodes.
- Webhook
- A web address another system calls the moment something happens, which starts a workflow.
- Hermes
- My own AI agent on a server, reachable from Telegram. It remembers, reads my notes and runs jobs unasked.
- Telegram
- The chat app I use to message Hermes and to get alerts.
- Knowledge base
- My searchable library of notes and documents, split into 1,277 short passages.
- Spec
- A short written brief: the outcome, the limits, the files and the checks, agreed before building.
- Acceptance criteria
- The checks a result must pass to count as done, written before the work starts.
- Claude Code
- The AI coding tool that builds: it reads and writes real files and runs commands.
- Subagent
- A helper Claude Code starts for one part of a task, so several parts are worked on at once.
- VS Code
- Visual Studio Code, the code editor I use to read what was written and take over by hand.
- Codex
- A second AI model that reviews the work. A fresh reviewer catches what the author missed.
- Skill
- A reusable procedure an agent loads for one kind of job, like drafting an article or cutting a reel.
- Agent
- An AI assistant with one job. It picks its own tools and finishes the task.
- Router
- The step that reads a job and hands it to the right specialist.
- Approval gate
- A hard stop. Nothing reaches a client, a repo or a send button until I approve it.
- Handover doc
- The same short document with every delivery: what it does, what breaks it and who to call.
- Write back
- Putting the result onto the client’s contact record, so the next run knows what already happened.
- Commit
- A saved, labelled snapshot of code changes.
- Deploy
- Putting the new version live: Vercel for sites, Docker on my VPS for Hermes and the workers.
- Roll back
- Putting the last good version back up when a live check fails.
- VPS
- A private server I rent, where Hermes and my workers run.
- Fold back
- Turning a mistake into a written rule, so it does not happen twice.
- Fresh session
- Starting the next task in a clean AI session instead of a long, cluttered one.