↓ Follow the highlighted line
astack by Applification
Outcome first.
Proof built in.
Engineering playbooks, focused skills and feedback for coding agents. Install the plugin in Codex or Claude Code, set up project model roles, then describe the outcome. Your project owns its technology choices and checks.
The repository is the source of truth. This site explains the workflow without pinning it to a release snapshot.
01 / ROUTING
One entry point.
Different kinds of work.
astack selects a playbook from the outcome you want. Each playbook names an ordered sequence, the skills it uses and the model roles that contribute. The skill's playbook guide has the detail.
THE ROUTE MAP
One task. The right line.
Choose a real request to see its expected decision and trace the line it takes.
ROUTES + REAL REQUESTS
Choose a request
Browse by route or search the examples. The selected line will light up on this map.
START HERE$applification:astack <task> → choose the line that fits the outcome
START HERE$applification:astack <task>Choose a route from the intended outcome
Define behaviorResearch · Read affected consumers
Settle boundariesDesign · Compare open choices
Build in slicesImplementation · Use project tools
Challenge the diffReview · Run real checks
Deliver the changeLead · Resolve findings + PR
Reproduce symptomResearch · Trace supported cause
Repair the causeImplementation · Add useful regression
Challenge the diffReview · Rerun failing path
Deliver the fixLead · Before/after proof + PR
Pin behaviorResearch · Name the invariant
Simplify ownershipImplementation · Move and compare
Challenge the diffReview · Inspect affected boundaries
Deliver the changeLead · Equivalence proof + PR
Measure baselineResearch · Pick a real metric
Find bottleneckDesign · Choose supported cause
Make improvementImplementation · Keep change bounded
Challenge the diffReview · Compare same conditions
Deliver the changeLead · Measurements + PR
Gather evidenceResearch · Code, history or runtime
Separate factDesign · Compare options if asked
Answer the questionLead · Sources and uncertainty
Ends with an answer
Read full diffResearch · Trace affected consumers
Challenge the changeReview · Read-only findings + checks
Fix findingsImplementation · Only when authorized
Resolve and reportLead · Recheck; deliver if authorized
Inspect productResearch · Reuse working drivers
Repair capabilityImplementation · Map paths when useful
Challenge the diffReview · Drive, observe, clean up
Deliver the changeLead · Resolve findings + PR
Understand productResearch · Fixed choices + milestone
Compare open choicesDesign · Use product constraints
Build first milestoneImplementation · Smallest useful path
Challenge the diffReview · Run real artifact
Deliver the milestoneLead · Resolve findings + PR
The lead owns integration and delivery. Read-only work stops with an answer. New project builds a first milestone; model setup configures the agent separately.
On a narrow screen, scroll sideways to follow the lines.
The map summarizes the playbooks; their linked files own the procedures. The agent keeps skipped steps visible with reasons. See the routing examples ↗
02 / METHOD
Playbooks order the work.
Skills supply the expertise.
A playbook defines the task sequence and model roles. Focused skills explain how to do each contribution. The lead owns integration and proof.
Understand the outcome
Read project guidance, affected consumers and the real check commands. Reproduce a symptom or name the requested behavior.
Investigation playbook ↗Settle open choices
Compare consequential alternatives against product constraints. Honour fixed technologies and record why a step can be skipped.
Architecture skill ↗Build in useful slices
Use the project's selected tools and platform guidance. Run fast checks before dependent work.
Feature playbook ↗Prove the integrated result
Observe the actual outcome, challenge affected boundaries and resolve findings. Read-only work ends with its answer.
Review playbook ↗03 / OWNERSHIP
Reusable method.
Project-owned choices.
astack ships 25 skills and eight playbooks. Your project supplies model preferences, application technologies and enforcement commands.
In astack
skills/
astack/SKILL.md entry point
astack/playbooks/ eight task procedures
setup-astack/ model-role setup
architect/ verify/ … focused expertise
principle-*/ seven principles
lint/ package + resolver checks
loops/ four trigger prompts
site/ public explainerIn your project
AGENTS.md / CLAUDE.md working guidance
.astack/models.json model preferences
source + tests your chosen layout
lint + compiler settings your chosen rules
check commands your actual feedbackUse existing guidance. .astack/project.md and product-driving tools are optional. The plugin adds no application stack.
MODEL SETUP
Choose the models.
Save choices with the project.
$applification:setup-astack uses the live host catalog to save and check research, design, implementation and review choices in this project's committed .astack/models.json. Reruns preserve choices and record inheritance explicitly.
Claude and Codex through T3
Keep a single reviewer or configure a two-provider panel. An explicit task choice wins; otherwise the panel runs when configured, or the saved single reviewer runs. Both panel targets must resolve before dispatch. Unavailable targets and options fail with a diagnostic; no model is silently substituted.
Setup checks read-only contributions on newly selected providers. Retained task handles and returned identities establish which models ran.
The main agent owns integration
It owns scope, delegates contributions, adjudicates findings and verifies the result. Briefs include skill paths, project rules, exact revision, allowed actions, expected proof and return format. Settings validation alone does not prove execution.
Work can begin with inherited roles. Setup checks the saved file; the host owns parent-model selection, persistent modes and scheduling.
FOCUSED SKILLS
Read what the step needs.
Enforce what can be checked.
The entry point selects the playbook and relevant expertise. You can also call a focused skill directly.
Engineering work
Useful contributions
FEEDBACK
Principles guide judgment.
Project checks enforce rules.
Improve the agent's working environment with deterministic development checks. Application source and runtime behavior stay unchanged.
Improve the agent's environment
Ask $applification:astack to prevent a demonstrated mistake. It uses the feature playbook with correct to change project tooling, lint/compiler configuration, structural checks, verification tools, hooks and CI. Prove bad contributions are rejected and valid ones pass; wire the command into local verification and CI and record it in project guidance. Product defects remain findings for separate work.
You can also call $applification:correct directly.
astack package checks
bun install --frozen-lockfile
bun run checkChecks reject malformed metadata, missing assets, broken skill/playbook links, and invalid model targets or options. Bun is this repository's authoring tool.
LOOPS
The same playbooks,
on a trigger.
A loop is a prompt. Your host owns scheduling, repository access, credentials and run state. Installing the plugin starts no jobs.
WEEKLYlint-debt
Fix one diagnostic using project checks.
HOURLY OR ON PRpr-review
Review new heads with the configured review role.
DAILYbug-fix
Reproduce and repair one eligible bug.
ON A MENTIONmention
Run the playbook for the comment's request.
Copy the prompt into your project and pass it to the host with the plugin installed, for example claude -p < agent-prompts/lint-debt.md. Pins require a matching host catalog; unsupported contributions stop with a reported gap. Runs avoid duplicate work and never merge.
NEW PROJECT
Describe the product.
Run a first milestone.
New-project work is a playbook. It honours fixed technologies, compares open choices and verifies the smallest useful end-to-end path.
$applification:astack Build a release-notes CLI.
Fixed: Python and the standard library.
Open: module boundaries and config format.
Done: parse a real config and print the release summary.PROOF
Run the real thing.
Keep gaps visible.
Proof names a revision, environment, action and observed result. Package validation and saved model choices do not prove an agent delivered the requested outcome.
Use the project's actual check commands and running-product tools. For a pinned role, inspect the host's returned provider/model identity. The lead resolves findings and verifies the integrated result; read-only work ends with its answer. Retain concise evidence in the PR when publication is part of the assignment.
09 / GET STARTED
Put it to work
in your project.
The installable plugin is called Applification. astack's guidance stays in this repo; your project owns its commands, environments, and product decisions.
TRY IT IN CODEX
codex plugin marketplace add applification/astack
codex plugin add applification@applificationInstall once, then describe the outcome in your project.
01 / INSTALL
Install the plugin
Use the commands alongside for Codex. In Claude Code, run /plugin marketplace add applification/astack, then /plugin install applification@applification.
02 / SET UP
Check project model choices
Run $applification:setup-astack in Codex or /applification:setup-astack in Claude Code. Save checked roles and a single reviewer or panel. Reruns preserve choices; inheritance is explicit.
03 / USE
Describe the outcome
Give $applification:astack an engineering request and let it select a playbook. Claude Code uses /applification:astack. Invoke any focused skill or principle-* leaf directly when you already know the job.