Build Packs
Packs for designing, deciding, and shipping code — Sentinel, FeatureDelivery, Agents, DreamTeam, Thinking, Compliance.
When do you reach for which Build Pack? Sentinel guards the code you already have. FeatureDelivery ships the next thing into it. Agents, DreamTeam, and Thinking are the three decision tools — parallel workers, named experts, structured debate. Compliance is the legal gate between your product and a Brazilian audit. Six Packs, six jobs.
The 6 Build Packs
Each Pack owns one decision you'd otherwise make by hand.
Sentinel — the repo guardian that scans once, then enforces forever.
FeatureDelivery — the end-to-end pipeline that classifies, specs, gates, reviews, and ships.
Agents — custom one-shot workers composed from Traits + Voice + Specialization.
DreamTeam — 7 named specialists channeling Peep Laja, Joanna Wiebe, Katie Dill.
Thinking — 7 modes from first principles to adversarial red team.
Compliance — LGPD-first gap analysis, policy generation, audit prep.
Sentinel
The cold-start problem, solved once. Sentinel scans a fresh repo, discovers its conventions, writes them into the knowledge graph and CLAUDE.md, then guards them forever.
Two halves. Scan is the 30–60 second awakening that runs once. Guard is the continuous check that fires on every staged change. Both share the convention cache, so a Guard fix teaches Scan next time.
When you'd reach for it
You just cloned a repo you've never touched, or a teammate's PR feels off but you can't name why. Ask Sentinel to scan and the answer comes back as 14 discovered conventions, a CLAUDE.md that encodes them, and a Guard mode that will flag the next violation for you.
Last Wednesday Sentinel Guard caught that a refactor had silently dropped the parent-id auto-detect on 28 docs files before the branch ever hit CI. Convention was two weeks old. Cost of the catch: 8 seconds.
FeatureDelivery
One skill, three tiers, one pipeline: Classify → Spec → Plan Gate → Build → Review → Ship Gate → Ship. Classification is the whole trick — Simple skips ceremony, Complex gets a 3-round council debate, Medium sits between.
Where Sentinel watches the code you already have, FeatureDelivery owns the code you're about to add. Worktree spun up, architect/engineer/designer/pentester Council gates on anything non-trivial, and no gh pr create until the 10-point review passes.
On April 2, 2026 a rate-limiting feature went in as Medium: worktree spun up in 3 seconds, spec written in 40, one round of Council at the plan stage caught that the limiter would starve health checks, fix landed before a single line of gateway code was written. PR opened 11 minutes after the initial request.
Agents
Not a team. Not a Council. A composition system for one-shot parallel workers built from Base Traits + Voice + Specialization, fired through Task(subagent_type="general-purpose") and forgotten about until they return.
The distinction matters. DreamTeam gives you named experts with pre-loaded frameworks. Thinking's Council runs structured debate. Agents gives you raw parallel capacity — spin up three workers with distinct perspectives, ship them against a PR, a competitor list, or a research question, and collect independent reports without orchestration overhead.
When you'd reach for it
Any time the answer is "I need three unrelated opinions fast and I don't care if they disagree." Competitor sweeps, PR reviews from opposing angles, quick research fan-outs.
DreamTeam
Seven named specialists channeling real frameworks: Peep Laja (CRO), Joanna Wiebe (conversion copy), Katie Dill (Stripe-caliber design), Karri Saarinen (Linear precision), Emily Heyward (Red Antler brand), plus Motion and 3D experts on demand.
Default workflow is SectionReview: read one section, hand 3 core experts the component inventory, implement only changes where 2 of 3 independently converge, verify, move on. The Unanimous Agreement Principle is the reason DreamTeam ships instead of generating critique-shaped noise.
A March 2026 pass on the Studio hero returned 6 changes — a Bento-to-CTA-stack swap, three copy tightenings that took the subtitle from 24 words to 13, a Magnetic hover on the primary CTA, two cuts. Every change tied to a named expert's framework. All 6 shipped.
Thinking
One skill, seven modes, routed by what you ask for.
| Mode | Use it for |
|---|---|
| FirstPrinciples | Rebuilding a decision from scratch, challenging inherited assumptions |
| IterativeDepth | Going deep on a single question from multiple angles |
| BeCreative | Divergent brainstorming, tree-of-thoughts idea generation |
| Council | Multi-agent debate with weighted consensus |
| RedTeam | Adversarial critique, failure-mode enumeration |
| WorldThreatModel | Time-horizon investment tests against future scenarios |
| Science | Hypothesis-driven iteration with falsifiable tests |
RedTeam is the one you'll use most. Ask it to attack a roadmap and you get failure modes ranked by plausibility, the hidden assumptions behind each, and the specific experiments that would falsify the weak links.
Compliance
LGPD-first GRC for Brazilian startups selling B2B. The core move is cross-framework mapping — LGPD and SOC 2 overlap ~40%, so one control implementation often satisfies two frameworks, and Compliance knows which.
Eight sub-skills cover the full engagement: GapAnalysis, PolicyEngine, DataMapping, EvidenceCollector, ControlMapping, IncidentResponse, AuditPrep, ComplianceDashboard. Twelve CLI scanners underneath (Bearer, Prowler, Checkov, Gitleaks, TruffleHog, Semgrep, SSL Labs, and friends) produce evidence feeding every workflow.
When you'd reach for it
A prospect asks for your LGPD posture on a Tuesday and you need a defensible answer by Friday. QuickScan returns a red/yellow/green scorecard in 90 seconds; PolicyEngine generates the 10 required LGPD policies in Portuguese; AuditPrep assembles the evidence package.
On February 19, 2026 a QuickScan on a 14-month-old Next.js codebase found 3 exposed API keys in git history, a missing SPF record, and 7 PII fields in logs — all in 94 seconds, all mapped to the LGPD articles and SOC 2 controls they violated.
Next: Operate Packs — the Packs for running DOS in daily practice once your repo is guarded and your features are shipping.
Was this page helpful?