Skip to content

ARTICLE —

3 concrete use cases with ulk

May 23, 2026 12 min EN ulk use-cases claude-code demo

3 concrete use cases with ulk

Three scenarios. For each: the starting situation, the key command, the agent delegation chain, the deliverable obtained.

Use case #1 — From zero prompt to a deployed Astro blog (Instagram → blog)

Situation

No written brief. No preparation. A single sentence typed into Claude Code. Goal: a static Astro blog fed from an Instagram account, deployed on Vercel, with a design system generated from the account’s visual identity.

This is exactly the scenario from the video script (video/script-demo-complete.md).

The single prompt

/ulk:bruce I want to build a tool that transforms an Instagram account
into a deployable Astro blog. Brief, architecture, design system from
the account's visuals, code, audit, Vercel deployment. You drive everything.

Automatic agent chain (11 steps)

1. Bruce (25) — routing

Receives the prompt, invokes Godspeed, draws up the battle plan in 11 steps, asks for validation.

2. Godspeed (00) — diagnostic

Empty repo. Also checks available CLIs (gh, vercel, node, claude) and the presence of a local LLM (ollama → 2b3 will use it for commits, 0 Claude tokens).

3. Tony (50) — stack + architecture + timing

Short questionnaire: post volume, source (Meta Graph API vs scraping), regeneration, image hosting, multi-account?

Output: Astro 5 SSG + content collections + strict TypeScript + Meta Graph API + Tailwind v4 + GitHub Action cron daily. Timing: 4 phases × 0.5 day = 2 solo dev days. Handoff to Stark.

4. Stark (58) — design system from Instagram

Analyzes the Instagram account URL, extracts the visual language via the Hue skill. Produces design-model.yaml (palette, typography, spacing scale, radii), tokens.css, and 6 components previewed in static HTML before writing a single line of TypeScript.

Targets: PostCard, PostGrid, PostDetail, StoryStrip, HeaderProfile, FooterMinimal. Handoff to Shuri for documentation, Brique (01 frontend) for implementation.

5. Shuri (01) — spec + todo + sync

Picks up from Tony + Stark. Produces docs/spec.md (frontmatter + 9 sections), docs/todo.md (Now: 8 items, full Backlog), syncs to CLAUDE.md and README.md.

6. project-decomposer (35) — 12 self-contained prompts

Transforms the todo into docs/prompts.md: 12 self-contained Claude Code prompts. Each works even without prior context — insurance against drift.

7. Guided implementation

Sequential execution of the 12 prompts. Sub-agent Explore at each heavy dependency exploration (hygiene rule 3). Official plugin /feature-dev delegated for half the vitest tests.

8. 2b3 (08) — clean checkpoint

tsc strict, lint, vitest (14/14 green), secrets scan, todo → Done, memory captured to docs/_memory/architecture/insta-pipeline.md, commit message suggested by ollama gemma3:1b local (zero Claude tokens for the commit).

9. Sargeras (45) — 10-axis audit before production

Security 92, performance 88, tests 78 (sync.ts at 62% coverage, added to Backlog), CI/CD 80, a11y 86 (4 missing alts → Backlog). Not blocking for v1.

10. Killbill (56) — cost audit

Vercel for this project: $0 (free tier OK for 500 static posts). On the global account: $60/month of waste identified on other dormant projects. Dry-run mode by default.

11. Xavier (57) → deployment

Checks repo, accounts (GitHub: izo ✓, Vercel: math-drouet ⚠️ personal, intentional for demo), Meta Graph token (read-only ✓). Then vercel --prod — deployed in 38s.

Bonus: scheduling cloud routines

/ulk:routine

Recommended bundle: 2b3 daily 18h, Godspeed weekly Mon, Sargeras weekly Mon, Robocop on check_suite.completed (failed), Lovecraft daily 8h, Killbill monthly 1st, ED-209 monthly 15th, plus the GitHub Action cron insta-sync daily 06:00 UTC. Autonomous sessions without a local machine.

Deliverables obtained

  • Static Astro blog in production, real images from the Instagram account, Stark design.
  • Complete documentation: docs/spec.md, docs/todo.md, CLAUDE.md, README.md, docs/prompts.md, design-model.yaml, tokens.css.
  • Clean git repo with a readable commit (“feat(v1): Astro blog from Instagram account + Stark design + daily cron”).
  • 14 green tests.
  • Initial Sargeras report archived for traceability.
  • 8 active cloud routines.
  • Vault memory ready for day 2.

Actual duration: ~24 minutes of filming, ~2 equivalent dev days solo without ulk.

Use case #2 — Auditing an existing project

Situation

A recovered repo (legacy, taken over from another dev, or simply not visited for 6 months). Goal: understand the current state, what’s problematic, how much it costs.

Steps

1. Mapping via Bruce (25)

/ulk:bruce

Bruce calls Godspeed (00), which scans the repo and produces a report: stack, state (greenfield / in progress / abandoned / production), key files present/absent, apparent debt.

2. Reverse documentation via Strange (16)

If the doc is missing or outdated:

/ulk:strange

Reconstructs docs/spec.md from the code. Mode mode=prompt: reverse-engineering prompts that produced a given AI output.

For Obsidian-first documentation:

/ulk:lovecraft harmonize

Detects the documentation state, migrates, intelligently completes.

3. Complete audit via Sargeras (45)

/ulk:sargeras

10-axis report. Lists violations sorted by severity.

4. Dedicated security audit via ED-209 (52)

/ulk:ed209

Focus: exposed secrets, vulnerable dependencies, attack surface, HTTP headers, CORS, auth.

5. Cost audit via Killbill (56)

/ulk:killbill

Vercel + GitHub + Neon audit. Dry-run mode by default. Quantified savings plan.

6. Context/token audit via Context Audit (55)

/context-audit

Context health score 0-100. Detects: unnecessary MCPs, bloated CLAUDE.md, over-sized skills, missing settings.

7. Synthesis via Blackemperor (18) audit mode

/ulk:blackemperor

Mode audit: aggregates Sargeras + ED-209 + Killbill + Context Audit outputs into a single report with priorities.

8. Action plan via Shuri (01)

Shuri picks up the recommendations and pushes them into docs/todo.md with the category “audit-2026-04”.

Deliverables obtained

  • Global project state report.
  • Prioritized fix list (security first, then costs, then tech debt).
  • Quantified cloud savings plan.
  • Updated docs/todo.md.

Use case #3 — Daily workflow

Situation

An active project, worked on every day. Goal: maximize productivity without saturating the context.

Morning routine

1. Startup: Xavier (57)

/ulk:xavier

Checks: right repo, right account, last git state, last session. (With the --with-xavier-hook hook, this is automatic at startup.)

2. Surface memory: Lovecraft (47)

/ulk:lovecraft memory surface

Displays a read-only summary of decisions from previous sessions. Auto-integrated in Godspeed’s phase 1.5.

3. Project pulse: Godspeed (00)

/ulk:godspeed

Current repo state diagnostic. Which tests are red, what was the last commit, where are we in the todo.

Development routine

4. One task per session

A single task from todo.md. No unrelated chaining (Rule 2: /clear between tasks).

5. Exploration via sub-agent (Rule 3)

To understand a module, don’t explore in main:

"Launch a sub-agent to summarize how the auth module works"

The sub-agent returns only the summary. Main context preserved.

6. Proactive /compact at 50-60% (Rule 4)

When the session approaches 50%:

/compact Preserve: architecture decision (option B), files in progress
(src/auth.ts), current bug (#A042). Forget abandoned approaches.

7. Error? Robocop (11)

/ulk:robocop

On broken build or red test.

End-of-session routine

8. Checkpoint via 2b3 (08)

/ulk:2b3

→ commit, todo updated, memory captured, clean git state.

9. /clear (Rule 2)

After 2b3, /clear for the next task.

Cloud routines (without a local machine)

Autonomous scheduled sessions via the Routine (53) agent. Recommended triggers (see CLAUDE.md section “Cloud Routines”):

  • 2b3 daily 18h — auto end-of-day checkpoint.
  • Godspeed weekly — weekly digest.
  • Robocop / CI Guard check_suite.completed (failed) — auto-fix red CI.
  • Sargeras weekly Monday — weekly audit.
  • Lovecraft daily 8h — surface memory before the day starts.
  • ED-209 monthly — monthly security audit.

Deliverables obtained (per day)

  • Todo tasks delivered.
  • Clean commits.
  • Memory up to date.
  • No context drift (the 4 rules hold).

Summary: who to call when

SituationAgent
I don’t know where to startBruce (25)
Vague brief → architecture + stackTony (50)
Visual identity / design systemStark (58)
Doc to create / synchronizeShuri (01)
Doc to reconstruct from codeStrange (16) or Lovecraft (47)
Complete auditSargeras (45)
Security auditED-209 (52)
Cloud cost auditKillbill (56)
Context/token auditContext Audit (55)
Broken build / red testRobocop (11)
End of session / commit2b3 (08)
Context verificationXavier (57)
Memory between sessionsLovecraft (47)
Schedule cloud routinesRoutine (53)

What’s next

Video: full end-to-end demo of a use case from scratch. See video/script-demo-complete.md.