Full-cycle development lifecycle
Why is itradioactive?
Because it decays. Every output is grounded in evidence — git diffs, test results, code reads. Nothing is left to imagination. When the agent ships code through radioactive, every claim has a citation, every fix references a finding, every decision is logged.
The premise
Everything validated.
Everything grounded.
The name is a joke with a serious point. In physics, radioactive elements decay — they lose energy by emitting radiation, transforming into something more stable. Radioactive the workflow does the same thing with code: raw, ungrounded, speculative output decays through 11 phases of validation until what remains is stable, tested, reviewed, and documented.
The deeper premise: as agents get smarter, the temptation is to let them run longer without verification. Radioactive resists this. Every phase is a gate. The agent does not proceed on assumptions — it stops, asks, verifies, and only advances with evidence. The mental model of the agent evolves with each cycle because Phase 10 captures what was learned into permanent contracts.
The tendency is always to evolve the agent's capacity. Not by giving it more freedom, but by giving it better structure. Each radioactive run produces institutional memory that makes the next run smarter — not because the model changed, but because the contracts did.
The pipeline
11 phases.
11 gates.
Need deeper discovery? The embed-aihero-radioactive variant expands this to 14 phases with the AI Hero methodology embedded.
CLASSIFY & INIT
Load the project skill, classify the task, extract constraints. If no skill exists, scaffold one via /workflow-blueprint init.
chains → /workflow-blueprintDISCOVER
Ask 3-5 Socratic questions. Wait for answers. No assumptions. The user defines the scope, not the agent.
chains → /brainstormingUX DESIGNconditional
If UI is involved: generate design tokens, color palette, typography scale, spacing system, component specs.
chains → /ui-ux-pro-maxPLAN
Break the problem into ordered tasks with file targets and dependencies. Write docs/tasks/task.<reference>.md. Gate on explicit user approval.
chains → /plan-writingBLUEPRINT
Transform the approved plan into a reusable workflow contract. Wire it into the project routing matrix.
chains → /plan-to-blueprintEXECUTE
Implement tasks sequentially. Read before editing. Minimal, targeted changes. Track progress in the task spec.
TESTS
Run the project test suite. Write tests for all changed code. Iterate until 100% pass. No exceptions.
REVIEW
Run thermo-nuclear code quality review on all modified files. Severity-ranked findings: Blocker, High, Medium, Low.
chains → /thermo-nuclear-code-quality-reviewFIX LOOP
Apply evidence-based fixes for Blocker and High findings. Run build. Re-review. Repeat up to 3 times until clean.
chains → /thermo-fixBLUEPRINTS
Capture discoveries into workflow contracts. Update project skills. The agent learned something — make it permanent.
chains → /workflow-blueprintCHANGELOG
Generate customer-facing release notes from git commits. Group by Features, Fixes, Improvements. Ship the story.
chains → /changelog-generatorThe variant
embed-aihero-
radioactive.
For complex features — new domain models, unfamiliar territory, high-risk logic — radioactive expands to 14 phases embedding six Matt Pocock / AI Hero skills into the pipeline. Discovery gets deeper: grilling instead of brainstorming, domain modeling with mandatory UUIDv7 keys, technical research spikes, and prototype validation before any spec is written.
Phase 1 always runs the primary project skill first for classification and constraint extraction — then hands off to /wayfinder for codebase orientation.
View full contract →/wayfinder
Phase 1Codebase orientation: entry points, data flows, route handlers, subsystem boundaries. Produces wayfinder.md.
/grilling
Phase 2Deep Socratic interrogation with trade-offs (Pros / Cons / Recommended). Produces decision-log.md.
/domain-modeling
Phase 3Aggregate Roots, Entities, Value Objects. Mandatory UUIDv7 PKs. Mermaid state + ER diagrams. Produces domain-model.md.
/research
Phase 4Technical spikes in scratch/ to resolve unknowns: dependency behavior, API constraints, performance. Produces research-spike.md.
/prototype
Phase 5Conditional POC for high-risk UI/algorithmic logic. Smoke-tested, recorded in prototype-report.md.
/to-spec
Phase 7Synthesizes all upstream evidence into docs/tasks/task.<reference>.md — numbered steps, file targets, verification commands, rollback playbooks.
11 → 14: the six AI Hero skills slot in as Phases 1–5 and 7 (to-spec). CLASSIFY & INIT becomes CLASSIFY & WAYFIND, DISCOVER becomes GRILLING, PLAN becomes TO-SPEC — and three new phases appear: DOMAIN MODELING, TECHNICAL RESEARCH, PROTOTYPE/SPIKE.
What it produces
Artifacts, not promises.
task spec
Executed task breakdown at docs/tasks/task.<reference>.md. Every task marked [x] or [/].
walkthrough.md
Decision log, UX specs, review tables, fix history. The full paper trail.
Workflow contract
Reusable SKILL.md for the feature. Next time, the agent loads this instead of starting from scratch.
Updated skills
Phase 10 captures new patterns into existing project skills. The system learns.
Git commit
Clean commit on local branch. Atomic, descriptive message.
Changelog
Customer-facing release notes grouped by Features, Fixes, Improvements.
Run it once.
The system learns forever.
Each radioactive cycle produces institutional memory. The agent that runs radioactive the 10th time is not the same agent that ran it the 1st time — not because the model changed, but because the contracts evolved.
View full contract →