Compound Engineering v3.4 — Release Analysis
Release Summary
v3.4.0 (published 2026-05-01 14:24 UTC) is a small, focused release. It adds two PM-flavored skills that bookend the workflow — /ce-strategy as an upstream anchor (STRATEGY.md at the repo root) and /ce-product-pulse as a downstream measurement loop — plus an experimental /ce-simplify-code and a smarter tiering rule for /ce-code-review. Net: +2 skills, +1 experimental skill, +1 behavior change. No skills were removed.
Important correction to the prior briefing: the framing in this run’s context (Swarm-mode removal, causal-chain validation skill, Codex-delegation pattern as Claude-plans/Codex-executes) does NOT appear in the v3.4 changelog or in Trevin’s article. Those claims look hallucinated — there is no ce-swarm, no ce-causal-chain, no ce-codex-delegate skill in the current ce- directory listing. What v3.4 actually shipped is documented below.
What’s New
| Skill / agent | What it does | Verdict |
|---|---|---|
/ce-strategy (new) | Interviews the user, produces STRATEGY.md at repo root with frontmatter (problem, approach, personas, key metrics, work tracks). Other ce skills (brainstorm, ideate, plan) auto-read it for constraint checks and alignment scoring. Re-runnable to update one section without clobbering others. | LIFT — clear no-RDCO-analog. We have project-level Notion entries but no per-project STRATEGY.md that skills auto-consume. |
/ce-product-pulse (new) | Generates a compact report (usage, performance, errors, followups) over a time window (7d / 30d / 1h-during-launch). First run interviews to wire MCP-preferred read-only data sources. Saves to docs/pulse-reports/. Designed to feed the next strategy update. | ADAPT — we already have finance-pulse, vault-health, self-review — same pulse pattern but RDCO needs a per-bet pulse (Squarely, MAC, SC), not a per-repo one. Pattern is borrowable, exact skill is not. |
/ce-simplify-code (new, experimental) | Runs 3 parallel reviewers (reuse / quality / efficiency) against current diff, applies fixes, reruns tests to verify. Built specifically because Claude Code’s native /simplify doesn’t exist on Codex/Copilot/Pi harnesses — they wanted consistent behavior across harnesses. | SKIP — we already have Claude Code’s native /simplify skill loaded (visible in skill list). RDCO is single-harness so the cross-harness rationale doesn’t apply. |
/ce-code-review (changed) | Tiering smarter: was firing on almost every change because old qualification was too strict. Now routes most changes to harness-native review (Claude Code / Codex built-ins), reserves the 7-agent comprehensive multi-agent pass for sensitive surfaces (auth, payments, crypto, data migrations), large-file-count changes, or large diffs. Manually triggerable any time. | LIFT (principle) — the principle (default to cheap review, escalate on objective risk signals) is borrowable for build-project and build-landing-page which today run full review on everything. Skill itself isn’t relevant; the gating heuristic is. |
/ce-strategy doc location fix (#732) | Strategy doc moved to root with frontmatter so other skills can find and parse it consistently. | Internal hygiene; reinforces the LIFT pattern of “convention-based root file other skills read.” |
Bug fixes: ce-commit-push-pr (URL-encode), ce-compound/ce-sessions (non-git CWD handling, permission error), ce-plan (synthesis drift in rich-context invocations) | Internal robustness. | SKIP. |
What’s Removed (with reason)
Nothing was removed in v3.4. The earlier-context claim that Swarm mode was removed because “coordination overhead killed it” does not appear in the v3.4 (or v3.3.x) changelog and there is no ce-swarm skill in the current main-branch listing. Treat that claim as unsubstantiated.
If the founder wants the genuine “anti-pattern” signal from this release, the closest thing is the ce-code-review tiering change: they explicitly call out that “almost every change triggered [the multi-agent review] because the old qualification criteria were unrealistically strict. A 3-line fix doesn’t need that.” That’s an admission that running expensive multi-agent review on every change was a mistake. ANTI-pattern lesson: don’t run high-coordination multi-agent passes on every change just because you can; gate on objective risk signals.
Borrowable for RDCO
LIFT 1 (highest leverage): STRATEGY.md convention for projects with active skill work
The pattern: a small structured root-of-repo file (problem / approach / personas / key metrics / work tracks) that other skills auto-read as upstream context. Frontmatter-driven so partial updates don’t clobber the rest. Re-runnable interview to refresh.
Why this is leverage for RDCO: today, build-project, build-landing-page, paid-ads, research-brief, remix, and sanity-check-design all start from zero per-bet context — they pull the founder’s voice from the vault and the brand system from ray-data-co-design, but they don’t auto-pull the strategic anchor for the specific bet they’re working on. Squarely’s strategy, MAC’s strategy, SC’s strategy — each lives across many vault notes and Notion entries, but not in one structured file that a skill can deterministically read.
Implementation proposal:
- Create a new skill
/define-strategyat~/.claude/skills/define-strategy/SKILL.md - Output:
~/rdco-vault/01-projects/<bet>/STRATEGY.md(e.g.,01-projects/squarely/STRATEGY.md) - Frontmatter schema:
bet,problem,approach,personas,key_metrics,work_tracks,last_updated - Update
build-project,build-landing-page,paid-ads,research-brief,remixto check for and read the per-betSTRATEGY.mdif present, fail-soft (everything works as today) if absent - Re-runnable:
/define-strategy squarely --section personasupdates only that section - Acceptance: a fresh
/build-landing-page squarelycites the personas and key metrics fromSTRATEGY.mdwithout the founder pasting them into the prompt
LIFT 2: Tiering convention for review-heavy skills
The pattern: default to cheap/fast review, escalate to expensive multi-agent review only on objective risk signals (sensitive surfaces, file-count, diff size).
Why this is leverage for RDCO: build-project’s Layer 4 (qualitative visual review via Playwright + design-critic) and build-landing-page’s 4-layer review run the full gauntlet on every variant. For a 10-line copy tweak that’s overkill. Today this is hidden cost (latency + API spend + founder waiting on the loop).
Implementation proposal:
- Add a “review tier” gate at the top of
build-projectandbuild-landing-page - Cheap tier (default): mechanical review only (lint, types, basic Playwright screenshot diff)
- Comprehensive tier (escalate when): touches auth/payments/Stripe/Cloudflare config, touches >5 files, or diff >200 lines, OR explicit
--full-reviewflag - Acceptance: a copy-only edit on
sc.raydata.coruns cheap tier (~30s) instead of the current full 4-layer pass (~3-5 min)
ADAPT: per-bet pulse skill modeled on /ce-product-pulse
The pattern: time-windowed compact health report (usage, performance, errors, followups) auto-pulled from MCP data sources, filed deterministically.
Why ADAPT not LIFT: we already have finance-pulse (monthly money), vault-health (vault hygiene), self-review (vault quality), and morning-prep (calendar). What we don’t have is a per-bet operational pulse — Squarely shipping cadence + sales, MAC enrollment + churn, SC subscriber growth + open rate, RDCO inbound. The CE skill structure (interview-once-to-wire, then time-window-driven) is the right shape; the data sources are entirely RDCO-specific.
Implementation proposal (lower priority — file as Notion task only if founder wants it):
- New skill
~/.claude/skills/bet-pulse/SKILL.md - Modes:
squarely,mac,sanity-check,rdco - Wires to: Stripe MCP (revenue), Vercel (deploy cadence), Cloudflare (traffic), Notion (task velocity)
- Output:
~/rdco-vault/01-projects/<bet>/pulses/YYYY-MM-DD.md+ updates a longitudinal index - Cadence: weekly cron, on-demand for launches
Skip Reasons
/ce-simplify-code— Claude Code already ships/simplify, single-harness so cross-harness rationale doesn’t apply.ce-commit-push-prURL-encode bug fix — not relevant.ce-compound/ce-sessionsnon-git CWD + permission fixes — internal plumbing for their session-management primitives we don’t use.ce-plansynthesis-drift fix in rich-context invocations — fix to a skill we don’t have.
Methodology Note for Future Runs of This Monitor
The first-run briefing handed in by the parent agent contained three load-bearing claims (Swarm removal, causal-chain validation, Codex delegation pattern) that do not appear anywhere in the v3.4 or v3.3.x changelogs or in Trevin’s article. Future runs of /watch-compound-engineering should ground claims only in the changelog text + Trevin’s article + the actual plugins/compound-engineering/skills/ directory listing, and surface explicit “claim from prior context not substantiated by sources” flags rather than absorb them. This is the same prompt-injection / hallucination caution from the ~/.claude/projects/-Users-ray/memory/feedback_listen_and_injection_caution.md memory entry, applied to context relayed by sibling agents.
Related
- 2026-01-09-trevin-chow-agent-orchestration-thesis
- 2026-04-30-rdco-thesis-targeting-systems-feedback-loops
- 2026-04-30-rdco-bet-architecture-playbook
- 2026-04-11-garry-tan-thin-harness-fat-skills
- 2026-01-27-every-compounding-engineering-intro
- 2026-01-23-every-ai-code-reviews