A delivery system for feature owners

The problem

A feature leader in a matrix organisation has responsibility without a reporting line. The only thing holding a feature together is context — and context is exactly what decays. Six weeks after a decision nobody remembers who was in the room or what the alternative was. Status lives in a chat thread. Risks are remembered right up until the moment they land.

At small scale you carry it. At five active features plus a dozen in support, carrying it is how things start slipping quietly.

What I built

Not a note-taking folder and not an "AI assistant". A delivery contour for a feature — from the brief through to post-release support — put together over a few months and then handed to the team as a method.

The feature contour

• Spec — scenarios, requirements, acceptance criteria, kept next to the feature rather than in a separate document nobody opens
• Decision log — every decision with its date, participants and reason; the answer to "why did we do it this way" survives the people who were there
• Status with history — a snapshot taken from the tracker every morning, append-only. Yesterday's status can't be quietly improved today
• Risk register — owner, review date, automatic resurfacing. A closed risk isn't deleted, it's marked with what actually happened
• Delivery forecast — reads closing speed and scope growth together, and calibrates on the feature's own history rather than on a wish
• Feedback synthesis — demo and retro notes reduced to what changes the plan, with call transcripts kept as primary sources
• Epic breakdown — epics sliced into stories and subtasks with assignment, following the team's own conventions
• Relationship map — entities, people and decisions as a navigable graph; every node opens the thing it describes

Around it

• Documentation and tickets updated automatically after a discussion, back into the tools the team already uses
• A release-day pulse posted to the team channels every fifteen minutes
• Analytics events written to the warehouse, so the system's own usage is measurable
• A separate release dashboard for leads and product managers — feature status, release date, blockers across two teams, rebuilt automatically at 9:00 every morning

Where the model is, and isn't

The language model writes text — spec drafts, decision summaries, stories. Statuses, history and forecasts are computed by scripts, with no model involved. Numbers are deterministic and reproducible: run it twice, get the same answer, and be able to show where it came from.

That line matters more than anything else in the list: a delivery tool that sometimes invents a date isn't a delivery tool. It runs through a corporate LLM proxy, so nothing leaves the perimeter.

What it changed

• Scope growth became visible instead of anecdotal — on one epic the system tracked the scope moving from 13 to 39 tasks, with the date each addition arrived
• Status stopped being a meeting. The answer exists before anyone asks for it
• Risks get reviewed because they resurface on their own, not because somebody remembered
• Leads read a dashboard instead of collecting statuses from people

What happened next

I demoed it to the department in September 2026. The ask afterwards was to make it available beyond my own team: published so it's findable in search, kept in git, with its own channel and a regular digest.

That's the signal I'd trust over the reaction in the room — a request to turn something into shared infrastructure means people expect to depend on it.

Notes for anyone building the same thing

• Start with the questions, not the tooling. Three recurring questions justified the whole system; everything that didn't answer one of them got cut.
• Append-only beats editable. A history you can rewrite is a history nobody trusts, including you.
• Give the model the writing, keep the arithmetic. The credibility of the whole thing rests on the numbers being boring and correct.
• Ship the method, not the favour. A template turns one person's system into a team practice. Setting it up for people turns you into their support desk.