Prepared for Tixel · April 2026
evolved.
day one.
Infrastructure
by week six.
You're not starting from zero: CLAUDE.md, Figma and Context7 MCPs, and Playwright work are already in play. The brief is to turn that experimentation into a shared baseline the whole team can compound on: one tool chain, one shared way of working.
Two days on-site up front so outcomes land immediately. Then four weeks embedded in your repos, tickets, and Slack, hardening the toolchain, widening loops to every seat, and leaving behind the practice muscle to keep the compounding going.
any one client.
Every team.
First wins come from coding speed.
Individuals get faster quickly. It's real, visible in days. But it's only the first layer, and the durable leverage lives elsewhere.
The next bottlenecks move elsewhere.
Planning quality, business logic, review, release confidence, and shared context become the real constraints, usually within weeks of the first speed bump.
Tool adoption alone isn't enough.
One live example: a seven-person team Gumbo worked with saw AI-assisted output rise materially on a roughly constant team size. The next-order work was review drag, release safety, and context retention.
The workshop is only the start.
Real leverage comes from follow-through: weekly operating reviews, tool and process refinement, and installing controls before exceptions become the operating system. This proposal is structured for that, not a one-off training.
on-site kickoff,
four weeks
embedded.
Kickoff & first loop
Repo audit, toolchain install, three parallel tracks. Converge to ship one real ticket. Milestone 01.
Agents, evals, playbook v1
Two agents live on trunk, eval harness in CI, playbook v1 merged. Milestones 02 · 03 · 04.
Soak week
Team runs the loops on real tickets. Async support in Slack. We watch what sticks and what stalls.
Harden & extend
Agents iterated on real PRs. Eval cases grown. Weekly paired sessions with rotating engineers.
Every seat covered
Loop audit across all ten attendees. PM & designer loops hardened with observed usage. Milestone 05.
Handover
Playbook v2, written retro against every milestone, 30/60/90 plan with named Tixel owners. Milestone 06.
On-site kickoff,
then embedded.
The core coding loop.
Claude Code signed in for all ten attendees, with a shared .claude/settings.json committed to main, MCP servers wired to GitHub, Jira, Figma, and Slack, and a starter set of team slash-commands. Installed Day 1; tuned against real diffs across weeks 3–4.
PR-review & ticket-to-spec agents.
Two agent scaffolds arrive pre-built, wired on Day 2 to MCP servers for GitHub, Jira, and Slack via a scoped Tixel service account. Post-workshop: iterated on real PRs, weeks 3–5.
The harness that keeps it honest.
An eval runner in GitHub Actions, 8–12 cases seeded from real Tixel tickets, a review-gate workflow template, and lightweight metrics logging. Grown to 20+ cases across weeks 3–5.
The Tixel way, written down.
A living playbook in your repo, started in the on-site, grown weekly as loops settle and PR-review patterns surface new rules, and handed over at Week 6 with a named Tixel owner against every loop and tool.
team is already
moving through.
AI changes the bottleneck.
Faster code generation shifts the real work toward planning, business logic, review, release, and context. The constraint moves; the team needs to move with it.
Planning as engineering work.
PM, design, and engineering learn to turn rough intent into better specs, fewer assumptions, and faster validation loops, before a single line gets written.
Shared memory that compounds.
Playbooks, slash-commands, and writeback loops move out of someone's private terminal and into team infrastructure.
Risk, review, and release control planes.
Risk tiers, reviewer routing, evals, preview environments, and safer deploy paths, installed before exceptions become the operating system.
Humane AI-native org design.
Coherent roles, less review thrash, and fewer avoidable incidents across the team. Throughput is only one piece of the picture.
Follow-through that compounds.
Weekly operating reviews and 30/60/90 checkpoints keep the workshop from becoming a one-off burst of enthusiasm.
keep.
Every seat on Claude Code
All ten attendees signed into Claude Code, with a shared .claude/settings.json committed to main and MCP servers for GitHub, Jira, Figma, and Slack wired via a scoped Tixel service account. One interface, one shared muscle.
A Tixel playbook, in-repo
Your team's loops, rules, and prompt conventions written down in your repo. The single source of truth Claude Code reads on every diff, amended by PR rather than by Slack argument.
Reusable Claude Code skills
Packaged slash-commands for your highest-leverage loops: /review-pr for parallel multi-specialist review, /pr-readiness so first-pass AI output gets caught by the author instead of the reviewer, /amend-playbook when a rule emerges, /classify-pr-risk to tier PRs before review.
Claude review in GitHub Actions
Auto-review on every PR opened, @claude review on demand in comments, and a post-merge pass that proposes playbook amendments back to the repo from review-comment patterns.
Eval harness + PR risk classifier
8–12 evals seeded from real Tixel tickets in CI, plus a tiered risk classifier that routes each PR to the right reviewer (human, agent, or both) before a line gets read.
Handover plan + written retro
A written retro against every milestone, named Tixel owners per loop and tool, and three 30-minute check-ins booked at +30, +60, and +90 days, included in this fee.
Simple terms.
Ready when you are.
A countersigned copy returned by email is sufficient to lock dates and kick off pre-work. Start date TBD.