The nine jobs
Each job is one skill. Your assistant loads only the one your request needs.
Which job runs
| User said something like… | Job | Load this skill | Write this |
|---|---|---|---|
| “Build me an agent that…”, “connect X and Y”, “I want a digest / bot / suite” | Scope | scope-to-suite |
<project>/blueprint.md |
| “Implement this”, “make it work”, “write the code”, “write the LangGraph / Crew / SDK” | Implement | implement-agent-from-blueprint |
code + tests + a prompt file and eval set per LLM step (Quality card) |
| “Test it properly”, “try to break it”, “is it really safe?”, and automatically after Implement | Verify | verify-agent-suite |
verify-<date>.md (testers, fixes with regression tests, re-test, council summary, verdict) |
| “Make it run”, “make it run every morning”, “connect my real accounts”, “put it live”, “deploy it” | Run | run-agent-suite |
the schedule + RUNBOOK.md (handoff.md when a developer should own it) |
| “Document it”, “write the docs”, “a README / architecture doc / user guide for this” | Document | document-agent-suite |
shared/docs/ (agent cards, tools and permissions, operations, architecture, ADRs, user guide) |
| “It’s broken”, “the digest is wrong”, “this trace failed” | Debug | debug-failing-agent |
debug record + a regression case |
| “Review this system”, “we inherited this”, “is this safe to ship?” | Audit | audit-existing-agent-system |
audit report with evidence |
| “Raise this”, “raise it”, “lead standard”, “use the schemas” | Raise | raise-existing-agent |
lead-audit-<date>.md + upgrade-plan-<date>.md (Together lines if 2+ Fails) |
| “How do I use this pack?”, first message, unclear | Bootstrap | this skill | the route table above |
Start — using-agent-systems
Start every session. Pick one of the daily jobs, load that skill, and produce an artefact in the user’s project.
Steps
- Set the builder mode once (Vibe / Developer / Together, below). Then match the user line to the table (Scope / Implement / Verify / Run / Document / Debug / Audit / Raise / Bootstrap).
- Load only that child skill. Do not load pattern groups until it names them.
- Copy the closest worked example (table below) before inventing a shape.
- Paste the cards your job’s skill names, verbatim (each child skill lists its cards; all nine are below for reference). Research before any claim about an outside system; cite the source or mark it
UNVERIFIED. - Write the artefact in the user’s project. Stop when that file exists and the child skill’s acceptance bar is green.
Scope — scope-to-suite
Turn an informal prompt (“connect Slack, Gmail, and Jira; give me task updates”) into a Suite Blueprint — one file that later skills implement. This is the Scope job.
Steps
- Copy the closest worked example (
using-agent-systems, Worked examples;shared/examples/task-digest/for a digest) as the shape. Do not invent a new layout. - Write
<project>/blueprint.mdfrom the § 12 template. Fill all 12 decisions in § 2’s order: intent, scope boundary, data sources, integration research (with the MCP access card), archetypes (with Together lines), memory, tool allowlists, trigger, output contract, guardrails (with the human-gate card), done and eval pass bar, iteration plan. - Paste the human-gate card. Unsigned Send / Delete / money → Implement is later forbidden. The named signer approves through the approval channel; you record the
**Signed.**line (§ 11). - Name connectors. Prefer an existing MCP server. Do not invent Send.
- A blank cell is a stop, not a guess. Research the real API before naming a tool. Each § 12 cell follows
shared/schemas/cell-contract.md(type / refuse / check / replay). Adjectives and"the user"are refuse. - Stop when
blueprint.mdexists and every cell has an answer or an explicitunknown. No code in this skill.
Implement — implement-agent-from-blueprint
Turn a filled Suite Blueprint into code. This is the Implement job.
Steps
- If
<project>/blueprint.mdis missing or a cell is blank → stop; runscope-to-suite. Do not guess. - Open
shared/schemas/frameworks/INDEX.md. One framework only (plain Python, stdlib only, by default for a small suite; LangGraph if graph; CrewAI if crew; Claude Agent SDK if Anthropic harness; Mastra/Vercel if TS+MCP; n8n / OpenClaw skills if those substrates). - Map: each blueprint agent → one module; each allowlisted tool → one typed wrapper; output contract → Pydantic/Zod + renderer.
- Scaffold
src/main,src/state,src/agents/,src/tools/,tests/test_golden_path.py. No dashboard. No extra agent. - Implement Analyzers/Reviewers → Pipeline → Resolver. Orchestrator last and only if named.
- Every mutating tool has a dry-run. Unsigned Send/Delete/money blueprint → do not write code. Read agent ≠ write token.
- One fixture → one run → one assertion against the output contract. Fill one eval-case row (
id,input_ref,expected_tools,pass_bar,sad_path,pairing_id) fromprepare-eval-set.--success trueis not that assertion. Copy the wholeshared/examples/task-digest/folder when the job is that digest (langgraph/alone fails: its fixtures live in../outputsand../edges). - Live adapters (Step 5b): each typed wrapper gets a live adapter behind the same interface. A scheduled job cannot use the chat host’s MCP connectors, so it needs its own credentials (OAuth client or API token, read scope first). Run § 3 connects them.
- Before you say done, meet the Quality card (Step 6): static checks run and recorded, one test per edge fixture, a prompt file and an eval set with a numeric pass bar for each LLM step (run, not
eval: not run), and tests that hold each prompt to the code. One golden assertion is one row, not the bar.
Verify — verify-agent-suite
Before a suite goes live, try to break it the way a hostile user, a careless operator and a buyer would: fresh testers attack in parallel, fixers repair, fresh testers attack again, and every miss becomes a mechanical check. The builder’s own tests only prove what the builder thought of.
Steps
- Record the revision under test (
git rev-parse HEAD) and the builder mode. Make a checkpoint commit. - Say roughly what the run uses (§ 2, budget), write one brief per tester (§ 3) and start all seven at once as fresh sub-agents, each report-only (§ 2).
- Triage: dedupe, reproduce every SUSPECTED, assign severity (§ 4).
- Fix: one fixer per folder, each fix with a test shown failing first, then passing.
- Re-test wave: fresh testers on the fixed revision. Repeat up to the round budget (default 3).
- Retro: for each confirmed finding, name the Quality-card row that missed it and add the mechanical prevention (§ 5).
- Council (§ 6): the in-host council every time; the paid external council only on the user’s yes in that turn. Write
verify-<date>.mdwith the verdict. Report in plain words.
Run — run-agent-suite
Take a suite that Implement built and tested, and make it run for real: on a schedule, on the user’s own accounts, with alerts, a dry-run trial first, and a named person’s yes before anything goes live. This is the job that lets a vibe coder finish without a developer, and that hands a developer a clean file when one is needed.
Steps
- Preflight: run the tests and the Quality card checks yourself and show the result. Red tests → stop. A failed or unrun card row (an LLM step below its eval pass bar or with no eval run, no independent review) → the Quality gate above: dry-run only, nothing live. Unsigned blueprint on a suite that can Send, Delete or move money → stop.
- Pick where it runs (table below): the simplest place that meets the blueprint’s trigger. Research the current setup steps for that place (Research card).
- Connect each account one at a time, read scopes first, through the live adapters Implement built. Secrets go in the place the table names, never in chat, never in git, never by asking the user to edit a file.
- Run a dry-run trial: every Send, payment, filing or refund stays
pending_approvalfor the trial window the blueprint sets (default 7 days). Alerts on (Alert card). - Go live only on a named person’s yes (§ 5b), recorded with date and scope. One switch per mutating tool, each set on its own yes; one master off-switch rolls everything back. No open Critical or High finding at go-live (§ 5, open findings); the handoff (§ 6) is separate.
- Write
RUNBOOK.md. Hand over withhandoff.mdwhen the table at § 6 says a developer should own it.
Document — document-agent-suite
Turn a built suite (blueprint, code, tests, RUNBOOK.md, alerts, audit) into technical documentation that is true today: written by the assistant, checked line by line against the code, and split by who reads it. This is the Document job.
Steps
- Preflight. Run the tests yourself and keep the output. Red → stop. List the inputs you found and the ones that are missing (no
RUNBOOK.mdmeans “not run for real yet”, not “nothing to say”). - Pick the set (§ 2). Full set when the suite moves money, files with a regulator, or holds other people’s personal data beyond the names and emails of the user’s own contacts. Otherwise the minimum set.
- Build the fact inventory from the code, not the blueprint (§ 4): agents, tools and allowlists, write paths and gates, schemas, files written, commands, exit codes, alerts, hooks, tests. Where the code and the blueprint differ, raise it with the user as a finding. Do not document either version until it is settled.
- Write reference first (agent cards, tools and permissions), then operations, architecture, ADRs, and the user guide last.
- Run every command you show and paste the real output. End every section with a
Source:line (file, function or line, test name). - Check the docs against the code (§ 6 claim check). Mark anything you could not check
UNVERIFIED. In Vibe mode, the user reviews only the user guide; you do the rest. - Add the changelog line and the regeneration rule (§ 7). Stop when Verification passes.
Debug — debug-failing-agent
Go from “it failed” to a named failure, a first-wrong step, a fix, and a test that catches it next time.
Steps
- Pick one failing run. Dated → do not use evidence from after that date.
- Name the failure class (skipped gate, expired auth, tool not allowlisted, poisoned tool text, unchanged retry,
--success truetreated as a verifier). - Mark the first-wrong step. Do not start at the last symptom.
- Fix that step only. Same-session review of the patch is not isolated — use a fresh child or a developer; if neither is available, record
review: not isolatedand do not go live. - Write
<suite>/debug/<date>-<short-name>.mdand one regression case. - If the record leaves the machine, follow Trace-release (keep subject + outcome; strip procedure). No single run but they want lead-grade →
raise-existing-agent.
Audit — audit-existing-agent-system
Review a system that already exists (inherited, acquired, or “we built this last quarter”) and say what is missing, with evidence.
Steps
- Inventory agents, tools, gates, connectors, and what can Send.
- Cheap pass first: allowlist per role, write-scope on a read agent, Send without a named human. Those two last are Critical.
- No severity until the finding is confirmed. If this session wrote the system under audit, the review is not isolated — use a fresh child or a developer; if neither is available, record
review: not isolatedand do not go live. An inherited system needs no fresh child. - Contest only with a file that child could have read. Verbal facts escalate; do not pad the artefact.
- If the run is multi-stage: write each stage file before the next; do not overwrite a completed stage. SDK
successis not terminal. - Write
<suite>/audit-<date>.md(inventory, evidence, ranked fixes). Schema scorecard / lead standard →raise-existing-agent. Do not walk the catalogue here.
Raise — raise-existing-agent
Attach to an inspectable existing agent (a folder you can read) and write a lead-standard review plus a PM upgrade plan. This is the Raise job.
Steps
- Fail closed if there is no folder to read. This is not a black-box wrap.
- Fill the five PM intake rows. Paste them at the top of both output files.
- Score the 14-row lead scorecard. Every Fail cites a Check-path. Fail-if is mechanical (
shared/schemas/cell-contract.md). Same-session reviewer or “done” from a Stop hook is a Fail. - Load only the schema slice that Fail names. Never the full pattern library.
- Two or more Fails → compose card: one Together line per pair (A then B), never a hybrid template.
- Write
lead-audit-<date>.mdandupgrade-plan-<date>.md. Report-only. Do not apply patches. Pressure:shared/examples/task-digest/pressure/raise-apply-patches.md.
(c) 2026 DesignUp Strategy AZ LLC. All rights reserved. Agent Suite Builder is licensed one seat per licence, sold on Gumroad; these pages may be read and linked to, not copied or republished.