Documentation is a force multiplier for AI-led teams — it externalises context that models can consume, reduces onboarding friction, and makes decisions auditable when prompt strategies evolve.
Teaches through doing. Leads the reader step by step. The measure of success is the reader achieving something tangible.
Guides through a specific task. Assumes the reader knows what they want — they need the steps. Success = task completed.
Describes the system accurately. Written for readers who know what they're looking for. Consulted, not read from start to finish.
Builds understanding of why. Discusses concepts, context, history, and alternatives. Does not instruct.
Framework Matrix
Every document type mapped to its phase and documentation mode simultaneously.
Tutorial
Learning-oriented
How-To Guide
Task-oriented
Reference
Information-oriented
Explanation
Understanding-oriented
↑ Click any document chip to view full details
Pre-Sales Track · 9 Phases
Everything above this line assumes a contract already exists. It doesn't — not yet. The RFP & Bid Lifecycle is a separate track with its own owners (capture and bid management, not delivery), its own gate reviews (bid/no-bid and deal desk, not design review), and a hard external deadline the customer sets. It runs from qualifying the opportunity through award, debrief, and a structured handover — which is exactly where it connects into SDLC Phase 01: Discovery & Ideation below.
The Nine Phases
Color-Team Review Glossary
Reviews capture strategy and win themes before solutioning begins.
Reviews storyboards and the annotated outline before full writing starts.
Scores the complete draft exactly as the customer's evaluators would.
Reviews cost, price, and commercial risk before it's locked in.
Final executive review of the submission-ready package.
Post-submission lessons-learned review, win or lose.
Cross-Phase Principles
Compliant is the entry ticket. These eight principles cut across all nine phases and are the difference between a proposal that scores and one that wins.
Qualify hard, then commit fully
The cheapest bid to lose is the one you never start. But once past bid/no-bid, resource it to win — half-hearted bids cost almost as much as serious ones and win nothing.
Compliance is the entry ticket, not the win
A non-compliant proposal is rejected unread; a merely compliant one loses to a responsive one. Compliance matrix first, win themes second — but always both.
Write for the evaluator's job
Evaluators score dozens of documents against a rubric under time pressure. Mirror the RFP structure, answer first and elaborate second, make every scoring criterion findable in seconds.
Themes need proof
Every claim — "proven", "experienced", "innovative" — that isn't attached to a metric, a named project, or a verifiable artifact is filler that dilutes the claims that do have proof.
Price is designed, not discovered
Bottom-up cost and top-down price-to-win are built independently. The gap between them is closed by explicit decisions — descope, redesign, margin — never by quietly shrinking estimates.
Reviews only work when they can hurt
Pink, Red, and Gold teams need independent reviewers, the customer's real criteria, and enough calendar left to act on findings. A review that can't change the outcome is a ceremony.
Everything you say becomes the contract
Proposal text, clarification answers, orals claims, negotiation concessions — assume all of it is binding, and review it at that standard.
The bid isn't done until delivery accepts it
A win handed over as a document dump becomes delivery's problem and next year's reference-client risk. The handover pack is part of the bid, not an afterthought.
Where the bid becomes the project
Award, Debrief & Transition
Discovery & Ideation
The Sales-to-Delivery Handover Pack carries the solution, assumptions, risk register, and budget straight into the Opportunity Brief — the bid team's knowledge becomes the delivery team's starting point.
Communication Matrix
Who owns, contributes to, and stays informed on each document type.
| Document | Engineering | PM | Design | QA | Ops/SRE | Leadership | Legal/Compliance | End Users |
|---|---|---|---|---|---|---|---|---|
| Phase 1 — Discovery & Ideation | ||||||||
| Opportunity Brief | ○ | ● | ○ | — | — | ● | — | |
| User Research Notes | ○ | ● | ● | — | — | ○ | ||
| Competitive Analysis | — | ● | ○ | ○ | — | |||
| Stakeholder Map | — | ● | — | — | — | ○ | ○ | |
| Value Proposition | ○ | ● | ○ | ● | — | |||
| Phase 2 — Architecture & Design | ||||||||
| ADR | ● | ○ | — | ○ | ○ | — | ||
| Tech Design Doc | ● | ○ | — | ○ | — | |||
| System Design | ● | — | — | ○ | ● | — | ||
| API Contract | ● | — | ○ | ● | ○ | |||
| SLO Definition | ● | ○ | ○ | ● | ○ | — | ||
| Phase 3 — Planning | ||||||||
| PRD | ● | ● | ● | ● | — | ○ | — | |
| FSD | ● | ○ | ○ | ● | — | |||
| Risk Register | ○ | ● | — | ○ | ○ | ● | ● | |
| Definition of Done | ● | ○ | ○ | ● | — | — | ||
| Phase 4 — Development | ||||||||
| README | ● | — | ○ | ○ | ||||
| Changelog | ○ | ○ | — | — | — | — | — | ● |
| Test Plan | ○ | — | ● | — | ||||
| Coding Standards | ● | ○ | — | |||||
| Phase 5 — Review & Launch | ||||||||
| Release Notes | — | ● | — | ○ | ○ | ○ | — | ● |
| Runbook | ○ | — | ○ | ● | — | |||
| Incident Playbook | ○ | ○ | — | ● | ○ | — | ||
| Security Review | ○ | — | ○ | ● | ○ | ● | ||
| Launch Checklist | ● | ● | ○ | ● | ● | ○ | — | |
| Phase 6 — Handover & Maintenance | ||||||||
| Post-Mortem | ● | ● | — | ○ | ○ | ○ | — | |
| Handover Doc | ● | ○ | — | ○ | ○ | — | — | |
| DR Plan | ○ | — | ○ | ● | ○ | ● | ||
| Tech Debt Register | ● | ○ | — | — | ○ | |||
AI systems accumulate invisible decisions — which model version, which prompt strategy, which embedding dimensionality. ADRs make these auditable. When a model update breaks production behaviour, you need to know what was decided and why, not just what changed.
Model drift is not a bug — it is a property of production ML systems. Your runbook should include drift detection thresholds, retraining triggers, and rollback procedures. An on-call engineer should be able to execute the full response at 3am without Slack.
When AI capabilities are embedded in a product and the team that built them moves on, what gets lost is not the code — it is the intent. Which prompts are load-bearing? Which system messages cannot change without breaking downstream behaviour? Document it before you leave.