pray

pray

package manager for the language placed before inference

The same plaintext and markdown traveled between repositories — amkisko gems among them. That surrounding material is the language placed before inference: policies, templates, memories, formatting rules, workflow notes. It shapes attention, refusal, style, and action before any model answers. Prayfile lists it as dependency; pray resolves, locks, and renders AGENTS.md and other targets.

descend to start kiskolabs/pray

treat pre-inference input as a dependency

orientation

The same AGENTS.md fragments, checklists, and conventions still moved repo to repo — stale hunks, silent edits, weak lineage, rollback by hand. Prayfile records them in a manifest and lockfile; pray renders deterministically; verify checksums managed spans.

In prayer, orientation precedes information — ask, praise, confess, remember, surrender, repeat, listen. Breath, rhythm, silence, and repetition carry weight beside the words. pray takes that order: language revision-locked and rendered before uncertain response. Shared spans are prayers; pray confess records accept or reject; silence in the spec omits resolved hunks from a target while keeping them in the lock. language before inference.

Inference here means provider, tool, retrieval, field order, repetition, silence, situation — conditions under which output stays probabilistic. Prayfile holds the manifest, lockfile, and rendered files.

Checksums cover spans that affect output. Revisions pin instructions that bend behaviour. Render twice; bytes must match.

two ways in

terminal

resolve · lock · plan · apply · verify

Run pray locally or in CI. Resolution writes Prayfile.lock; apply writes rendered files into the repository.

  1. Install the CLI with Cargo.
  2. pray init scaffolds Prayfile and destinations.
  3. Declare compose, tree, and pray inputs (or migrate with pray format).
  4. pray install → pray plan → pray apply.
  5. pray verify and pray drift in CI.

tools

assistants · review bots · generators · wrappers

Tools read committed files. After apply, pinned text is in AGENTS.md, skills, and checklists; silence entries in the lock omit spans from that target.

  1. Wire shared packages in Prayfile (for example from amkisko/prayers).
  2. Render once; commit Prayfile.lock and inference-facing targets.
  3. Open the tree in your tool; on-disk targets carry pinned revisions.
  4. Update packages; review with pray plan before apply.
  5. Record feedback with pray confess.

start here

install

cargo install --git https://github.com/kiskolabs/pray --locked pray

Installs to ~/.cargo/bin. Add PATH if required.

scaffold

cd your-project
pray init
pray install
pray plan
pray apply

Writes Prayfile, Prayfile.lock, and rendered targets including AGENTS.md.

shared guidance

Example Prayfile using amkisko/prayers:

prayfile "1"

source "amkisko", git: "https://github.com/amkisko/prayers.git"

compose "AGENTS.md" do
  pray "amkisko/working-rules", "~> 2.0"
  pray "amkisko/ruby-conventions", "~> 1.0"
end

tree ".agents/skills" do
  pray "amkisko/engineering-audit", "~> 2.0"
end

Use rust-conventions or elixir-conventions on other stacks. Legacy target / agent forms remain valid; pray format rewrites them to this destination DSL.

daily

pray install    # resolve dependencies
pray plan       # preview rendered changes
pray apply      # write targets + refresh lock spans
pray format     # rewrite Prayfile to destination DSL
pray verify     # checksum managed spans
pray drift      # detect hand-edits and renderer drift
pray confess    # signed accept / reject feedback

project-specific

Shared prayers hold generic fragments. Project paths, credentials, and deployment detail belong in docs/, usr/docs/, and .agents/.

Three zones. Recipe — Prayfile, packages, lock — changes via pray add, pray update, pray apply. Human-owned .agents/ files stay editable; pray embeds them on render. Managed blocks between pray markers regenerate from lock, recipe, and compose-bound local paths.

recipe Prayfile · packages · lock

human-owned .agents/ · local embeds in compose

managed AGENTS.md blocks · .agents/skills

Put project-specific language in .agents/ and list it with pray inside compose.
Hand-edits inside managed spans surface as custom implementation in pray verify and pray drift.

compose "AGENTS.md" do
  pray ".agents/project.md"
  pray "amkisko/working-rules", "~> 2.0"
  pray ".agents/security.md"
end

tree ".agents/skills" do
  pray "amkisko/engineering-audit", "~> 2.0"
end

local ".agents/private.md", optional: true

Local paths re-embed on each render. Declaration order inside compose is render order. Declare silence to omit a shared fragment; the lock records it. For a project-only package: pray prayer init, then a path source in Prayfile. Whole-file destinations use pray "pkg", "~> 1.0", file: "SECURITY.md".

materialization

A manifest entry differs from bytes a tool loads. Truth: Prayfile.lock. Recovery: .pray/cache. Influence: rendered files opened at inference time.

truth Prayfile.lock

recovery .pray/cache

influence AGENTS.md · skills · tool targets

rendered files carry influence

amkisko/prayers

amkisko/prayers hosts shared fragments — conventions, skills, audits — referenced from your Prayfile.

working-rules
fragment
security
fragment
docs-conventions
fragment
dependency-issues
fragment
dependency-policy
fragment + skill
ruby-conventions
fragment
rust-conventions
fragment
elixir-conventions
fragment
minimal-implementation
fragment
engineering-audit
skill
changelog-update
skill
collaboration-workflow
fragment
getting-started
fragment
community-security
file
community-code-of-conduct
file
community-governance
file
community-contributing
file

full catalog

repositories

kiskolabs/pray amkisko/prayers