Take your team's AI from trial and error to documented agentic development
Many teams already use Claude Code, Copilot or Gemini, but every developer does it their own way: loose prompts, undocumented changes and costs nobody tracks. I install a process where every change starts from a spec, agents work by role, gates stop what is not ready and everything is recorded in the repo. It is the same method I use to build my own products.
What is included
- Assessment of how your team uses AI and a staged adoption plan
- SDD-Harness set up in new or existing repos, without touching application code
- Agents by role: orchestrator, functional, planner, architect, implementers and reviewer
- Automated gates in CI: no spec, no cycle; no cycle, no code
- Model, effort and token telemetry per task, with a cost viewer
- Team training and support through the first real cycles
A kit that adapts to the size of the change
The same records and validations, with just the right amount of ceremony for each case.
Full flow (team profile)
One role per document: brief, functional stories, planning and architecture before coding. For features with contracts or several teams involved.
Reduced flow
Refactors and technical changes without user stories: brief and tasks, with the same traceability.
Lite flow (solo profile)
A single actor with a one-file plan: ideal for individual developers or small changes. If the change touches a contract another system uses, it switches back to full automatically.
Traceable fixes
Hotfix, bugfix or improvement without waiting for a full cycle, but always recorded and reviewed at closing.
Built for several developers in the same repo
No ID collisions
Specs, tasks and fixes carry the author's GitHub username: two developers never generate the same identifier.
No merge conflicts in the context
Technical context is updated with additive fragments per cycle and a single actor consolidates them, so the most contested files never clash.
Everyone keeps their tool
The same kit works in Claude Code, GitHub Copilot and Gemini/Antigravity: nobody has to switch tools.
Shared team memory
Lessons learned are distilled in the repo and every agent reads them at the start: mistakes are not repeated.
Validation on every pull request
One command validates schemas, gates and records in CI: a PR with a broken process does not get merged.
Costs visible per agent
Every task records model, effort and tokens; the viewer shows what each cycle costs and where to adjust.
How to engage
Fixed-price audit
From USD 1,100 · fixed price
Assessment of the system and the team, prioritized risks and a concrete roadmap for the next stages.
Typical timeline: 1–2 weeks
Hourly consulting
From USD 45 · per hour
Design reviews, one-off decisions, pairing with your team or guidance through a migration.
Typical timeline: on demand
Dedicated sprint
From USD 3,600 · per 2-week sprint
Two weeks embedded in your team to build or migrate a specific piece, with clear deliverables.
Typical timeline: 2 weeks
Prices in USD based on a senior rate of USD 45–55 per hour. The final scope is defined in a free intro call.
Real cases
Flexibility — consultancy standard
The SDD harness is the standard methodology across its projects, together with a central template registry for AI agents I designed as tech lead.
YPF — several repositories in the ecosystem
Mandatory SDD in the operations back office, in funds disbursement with agents by role and in the platform catalog: no feature ships without a spec, a cycle and a review.
SDD-Harness and my own products
Open source CLI on npm (v0.15.1) with 8 agents and 19 skills. DisplayAds was built through 117 specs and 104 cycles with this method.
Frequently asked questions
What is Spec-Driven Development?
A methodology where every change starts from a registered specification, passes automated gates and closes with review and traceability. AI agents work within that process instead of flying blind.
Do we need to change our stack or existing code?
No. SDD-Harness installs on top of the repository without touching application code and works with any stack.
Does it work if each developer uses a different AI tool?
Yes. The kit exposes the same agents, skills and rules to Claude Code, GitHub Copilot and Gemini from a single source.
Isn't it too much process for small changes?
That is what the lite flow and the FIX GATE are for: records and validation stay, but with a single document or no questionnaire.
How long does adoption take?
The assessment and plan are delivered in one or two weeks; the pilot in one repository with the team, in one sprint. Then it extends to the rest of the repos.
Shall we talk?
In a free 30-minute call we go over your case and I propose where to start.