Skip to content

/method

// method for reliable AI-assisted delivery

// orientation

MAP — Read the principles at /about, see the artifacts at /work, then end up here: the operating loop that makes the work reviewable, testable, and safe to ship. /rpi, /crank, beads, ao are just the tools that run it. The idea underneath is plain: DevOps and SRE discipline, applied to AI work.

// the entry

How I enter depends on how well I understand the problem. The phases are the same either way. What changes is how much steering it needs before it can safely enter the loop.

high confidence
$ /rpi "add rate limiting"

Full auto. Six phases, zero prompts. The lifecycle carries its own validation.

medium confidence
$ /discovery → review → /crank

Run discovery, then review the plan and pre-mortem before any code. Steer the slices before cranking.

exploring
$ /brainstorm → /discovery → …

Don't even know the shape yet. Brainstorm first, then run discovery once the direction is clear.

// the methodology

Three moves: discovery, crank, validation. /rpi runs them as six phases: full auto or hand-driven.

$ /rpi "add rate limiting to the API"
// that's it. everything below happens automatically.
/researchexplore the codebase, gather context
/plandecompose into issues with dependencies
/pre-mortemmulti-model council validates the plan
/crankparallel workers execute in waves
/vibecouncil validates the implementation
/post-mortemextract learnings into the flywheel
.agents/research/
.agents/plans/
.agents/council/
commits + closes
.agents/vibe/
.agents/learnings/

Every phase writes to a persistent directory. Knowledge accumulates across sessions. Validation gates auto-retry on failure.

// the execution

/crank turns a plan into parallel waves of workers while keeping scope, dependencies, and proof visible.

$ /crank ps-kfi
// takes an epic ID; spawns workers per wave, coordinates dependencies
Wave 1 · 3 workers in parallel, no blockers
Wave 2 · 2 workers, after Wave 1 completes
Wave 3 · 1 worker, integration + tests

THE 40% RULE — Keep context under 40% utilization. Past that, in my runs, the model starts dropping instructions. Beads issues preserve state when you need to start fresh.

// the flywheel

Each session writes back what it learned, so the next one starts from it. Learnings persist in git instead of vanishing with the context window.

ao forgeextract learnings from session transcripts
ao poolstage for quality review, promote to knowledge base
ao injectload relevant learnings at session start

THE INSIGHT — It's not agent orchestration, it's context orchestration. Recency-weighted: recent learnings weigh higher, stale knowledge is evicted.

// the tracking

Three systems work together: beads for issues, .agents/ for knowledge, ao for the flywheel. The stack makes context survive execution.

beads

Git-backed issue tracking. Survives sessions, crashes, and context resets.

bd ready → unblocked
bd show <id> → details
bd close <id> → done
.agents/

Persistent memory. Research, plans, learnings, patterns, all in git.

research/ → exploration
plans/ → specs
learnings/ → insights
ao

Go CLI, ~70 commands for knowledge compounding. The memory half of AgentOps.

ao forge → extract
ao inject → load
ao status → health

// the validation

Multi-model councils validate before and after implementation: a panel, not a single reviewer.

/pre-mortem

4 judges review the plan before any code is written. Missing requirements, feasibility, scope, spec completeness. FAIL → re-plan automatically.

/vibe

6 judges review the code after implementation. Quality, security, architecture, complexity, UX. FAIL → re-crank automatically.

THE RATCHET — Progress is locked permanently. Code is merged, issues closed, learnings stored, all in git. The next session inherits the gain instead of re-earning it.

// the replays

Real commits from this repository showing the method in action.

7fd5e98May 2026

Add the /training page (coding-agent training lane) (#37)

Scoped a new lane, planned the page, shipped it behind review. Plan → implement → land.

1069b51May 2026

Feature the ADHD essay, lived opening, voice-pass /training (#38)

The validation gate in practice: a voice-pass on the copy before it shipped, not after.

f2b3008May 2026

Act on mixed-vendor post-mortem findings (#39)

A post-mortem across the site became a punch-list, then concrete fixes. The learning ratchets forward.

// the toolkit

AgentOps

Coding-agent plugin for any model. Go CLI + skills + marketplace kits.

/rpi
lifecycle
ao
flywheel
/council
validation
/crank
execution

This website, the CLI, the skills, the marketplace kits: all built with the same workflow they document. The repo is the proof. Every commit, session, and learning extraction happened through this loop.

// the bigger picture

This is the discipline that makes AI dependable for engineers, and it's what I teach teams at /training. I'm now translating it for people who aren't: /ai-partner brings the same context, boundaries, and safety to everyone, in plain language.