How We Built Sculpting Framework
AI coding agents made writing code fast. They did not fix ambiguous specs, late security vetoes, or mid-sprint scope drift. Here is how we designed a lightweight operating system for the work before commit #1 — and stress-tested it across seven simulations.
The Pre-Commit Bottleneck
If you work as a Product Engineer today, you have probably felt a strange shift: implementing a well-defined ticket takes hours instead of days, yet shipping a meaningful feature still drags on for weeks. The slowdown rarely happens in the editor. It happens before the first line of code is written.
A rough Slack request turns into a hurried implementation. Halfway through, Finance asks about EU VAT edge cases, Security blocks a third-party vendor that nobody vetted, or two tracks accidentally assume the same backend engineer has 200% capacity. Heavyweight enterprise processes try to solve this with 15-page RFCs and endless meetings; ad-hoc teams ignore it and pay with rework.
We built Sculpting Framework to occupy the missing middle: an open,
tool-agnostic set of five SKILL.md skills that any AI coding agent can run
alongside you to turn messy requests into aligned, implementation-ready vertical slices.
The Mental Model: Progressive Resolution
Think about how a sculptor carves stone. You never polish an eye socket before you have cut the rough silhouette of the head. If the proportion is wrong, every hour spent polishing is wasted. Software delivery fails for the exact same reason: teams jump into database schemas and UI states while the underlying problem and scope boundaries are still blurry.
Sculpting resolves ambiguity in four explicit cuts — Block → Rough Cut → Detail → Polish — where each stage has a crisp exit gate. And because a one-day tweak should never carry the same ceremony as a multi-team billing rewrite, an internal triage step routes tiny tasks (≤2 days, single service, single owner) straight through a zero-bureaucracy Fast-Track.
One non-negotiable rule runs across all four stages: zero meta-jargon in stakeholder messages. Your PM, designer, or finance lead should never read words like "Block gate", "Triage", or "SSoT." They only see crisp context, at most three decision-changing questions, and realistic options — never fake strawman options invented just to fill a comparison table.
Stress-Tested Across 7 Multi-Agent Simulations
Most process frameworks fail because they are written in a quiet room and only describe the happy path. Instead of guessing what rules we needed, we ran seven live multi-agent swarm simulations — complete with opinionated PMs, strict Security reviewers, delayed stakeholders, mid-flight pivots, and a final head-to-head A/B regression replay.
Every rule in v1.0 exists because a simulation broke without it:
- Sim 1 (Enterprise Billing, v0.1): Caught five structural failures — including deferring a risky tax-API spike to "Slice 0" and sneaking meta-jargon into emails. That single run gave us Templates 1–4, the Conditional Approval Lock, and Pre-Detail Spikes.
- Sim 2 & Sim 3 (SSO/RBAC & EU GDPR AI): Hardened Vendor/DPA gates, added the Spike Summary (Template 5) and Single-Owner Capacity Check, then converged cleanly at zero bugs and zero meetings in 12 async messages.
- Sim 4–6 (The Chaos Trio): We deliberately broke the happy path with a delayed PM, a 1-week P0 pause, an owner handoff, a 2-round security veto deadlock, and a 1-day CSV tweak. Those three stress tests produced the Minimal Recovery Protocol, the 25-minute Meeting Fallback with a tie-breaker, and the Fast-Track rule.
- Sim 7 (A/B Regression Replay, v1.0-rc): Re-ran Sim 1's exact billing scenario on the finished framework. All 5/5 original bugs vanished in 8 async messages and 0 meetings.
Surviving Real Chaos: 1 Card + 2 Rules
Sim 4 was the turning point. Real projects don't move in straight lines — a PM goes on vacation, a P0 outage pauses work for a week, or a lead engineer hands off ownership mid-flight. Instead of inventing a heavy governance layer, we distilled recovery down to one pinned card and two rules:
Dogfooding v1.0 → v1.0.1
The real test of any framework is using it on your own product. Right after freezing
v1.0, we installed the actual SKILL.md plugin and used it to build this
exact website — and then used it again to add this blog.
Dogfooding immediately surfaced four subtle rough edges that synthetic enterprise scenarios hadn't triggered. We fixed all four in v1.0.1:
- Conditional Pre-Detail Spikes: Pure static/UI work with no external APIs no longer needs a ceremonial spike — mark it N/A and keep moving.
- Solo-Owner Detail FYI Handoff: When one engineer owns the build and the sole decision maker already approved implementation at the end of Rough Cut, post the vertical slices as an FYI and go straight to Polish instead of adding an extra blocking approval round.
- Conditional Telemetry: Live metrics in the Polish Done Gate are explicitly conditional — never violate a privacy or no-tracker Red Line just to tick a box.
-
Manual-Mode Hygiene: Clarified where to store local project state
(
docs/sculpt-<slug>.md) and to keepsculpting.context.mdlocal in public repos.
The takeaway: You do not need a heavier process to get predictable engineering outcomes. You need the smallest set of gates that catch expensive mistakes while they are still cheap to change.