Rick is a local CLI for turning one-shot LLM prompts into judged, logged workflows.
It is built for work where the first AI answer is not good enough: writing, product docs, refactoring plans, code generation, and any task where you want candidates, a judge, and a trace you can inspect later.
Most AI tools optimize for a fast answer. Rick optimizes for a usable answer.
Instead of asking once and hoping, you define the task and Definition of Done, generate multiple candidates, judge them against the criteria, and only then produce the final artifact.
RESOLVE("TASK","DOD")>GEN(plan,3)>JUDGE>GEN(draft,3)>JUDGE>EDIT(strict)
That gives you:
- multiple attempts instead of one lucky sample
- explicit criteria through the Definition of Done
- a judge step that compares candidates
- readable
log.mdand machine-readablerun.json - a local terminal UI for repeated workflow work
- safety gates for file writes, shell verification, and context access
Rick is not an autonomous agent. It is a small workflow runner for getting better LLM output with less guessing.
Recommended install with uv:
uv tool install git+https://github.com/para1992/rick.gitOr with pipx:
pipx install git+https://github.com/para1992/rick.gitBoth install the same console commands:
rick --help
rick -iFor real OpenRouter runs, set an API key in your shell:
export OPENROUTER_API_KEY=...Or create a .env file in the directory where you run Rick:
OPENROUTER_API_KEY=...
OPENROUTER_MODEL_MEDIUM=google/gemini-2.0-flash-001From a cloned checkout, you can start from the example file:
cp .env.example .envLeave OPENROUTER_API_KEY empty to use the local mock client for CLI checks.
rick -iOr from a checkout without relying on the console script:
python3 -m rick_cli --interactivePaste a workflow, press Enter, and Rick will stream the run into the terminal UI. Press Ctrl-Q to exit.
rick 'RESOLVE("Write a concise product brief","Must explain the problem, target user, value, and concrete takeaway")>GEN(plan,3)>JUDGE>GEN(draft,3)>JUDGE>EDIT(strict)' --mode LOG_STEP --run-dir runs/product-brief --max-calls 10Outputs:
runs/product-brief/run.json
runs/product-brief/log.md
RESOLVE("TASK","DOD") define the task and Definition of Done
CONTEXT(path) add a local context file
GEN(artifact,n) generate n candidates for an artifact
JUDGE select the strongest candidate
EDIT(strict) produce a final edited artifact
OUTPUT_GLUE return the selected artifact directly
OUTPUT_AI_GLUE(strict) use the model to combine selected artifacts
UNFOLD(source,child,n) split a selected artifact into child artifacts
UNFOLD_JUDGE(source,child,n) split, generate candidates, and judge per unit
MATERIALIZE(path) write generated files under runs/
VERIFY("command") run a trusted shell check after materialization
If a workflow does not end with OUTPUT_GLUE, OUTPUT_AI_GLUE(...), EDIT(...), MATERIALIZE(...), or VERIFY(...), Rick appends OUTPUT_GLUE.
Rick uses seeders to make candidate generation more diverse and easier to audit.
For normal GEN(...) candidate generation, Rick uses a String Seed of Thought prompt. The model must first create a <random_string>...</random_string>, use it to make stochastic choices inside its private planning, and then return the final candidate JSON inside <answer>...</answer>.
Every generated candidate must include:
{
"seed": {
"random_string": "<random_string>A7kP2mQ9zR</random_string>",
"interpretation": "String manipulation shaped the candidate structure and tone."
}
}Rick stores that seed metadata in run.json and log.md, so you can see which candidate came from which seeded generation pass.
Where this lives in code:
rick_cli/prompts.pydefines the seed protocol prompt.rick_cli/engine.pyenables seeders by default for artifact generation.rick_cli/engine.pyvalidatescandidate.seed.random_stringandcandidate.seed.interpretation.rick_cli/llm.pyparses seeded model output from the<answer>...</answer>block.
Workspace prompt aliases can disable seeding for an artifact with "seed": false, but the default is "seed": true.
Separately, if you skip JUDGE between generation steps, Rick uses a runtime random seed to auto-select one candidate and logs that selection policy. That keeps implicit choices visible instead of silently picking the first answer.
Rick will create three competing product-brief plans, judge which plan best satisfies the Definition of Done, create three drafts from the winning plan, judge those drafts, and then edit the selected draft into one polished brief. Use this when you want a stronger positioning document than a single prompt usually gives you.
RESOLVE("Write a product brief for a CLI tool that turns one-shot prompts into judged LLM workflows","500-700 words, explain the problem, target user, core workflow, why judging beats one-shot prompting, and one concrete adoption path")>GEN(plan,3)>JUDGE>GEN(draft,3)>JUDGE>EDIT(strict)
Rick will first explore three possible ADR structures, select the strongest one, draft three full ADR candidates from that structure, select the best draft, and run a final strict edit. Use this when the output needs engineering shape: context, decision, alternatives, consequences, rollout, and risks.
RESOLVE("Write an ADR for using staged LLM workflows instead of direct chat prompts in a developer tool","Include context, decision, alternatives considered, consequences, rollout plan, and risks. Be specific and engineering-focused.")>GEN(plan,3)>JUDGE>GEN(draft,3)>JUDGE>EDIT(strict)
Rick will generate three story outlines, pick the one with the best tension and ending, draft three complete story versions from that outline, judge them against tone and safety constraints, and edit the winner into a final first-person post. Use this to test long-form voice, pacing, and non-English writing quality.
RESOLVE("Write a first-person Reddit story in Russian about a bad date where the narrator realizes the woman is being pulled into a cult","700-1000 words, conversational Russian, suspense, dark humor, no graphic violence, do not romanticize the cult, end with an uneasy aftertaste")>GEN(plan,3)>JUDGE>GEN(draft,3)>JUDGE>EDIT(strict)
Rick will create three angles for the post, judge the strongest angle, draft three short posts from it, judge those drafts for specificity and lack of fluff, and edit the winner into a final concise post. Use this when you want practical social content without generic AI copy.
RESOLVE("Write a short LinkedIn post for software developers about why judged multi-candidate workflows beat one-shot prompting","120-180 words, one concrete thesis, developer example, no marketing fluff, practical takeaway")>GEN(plan,3)>JUDGE>GEN(draft,3)>JUDGE>EDIT(strict)
Rick will generate three file manifests, judge which project structure best fits the browser-app requirements, expand the chosen manifest into implementation candidates per file/unit, write the selected files under runs/canvas-square, and run node --check against every generated JavaScript file. Use this for trusted local codegen where you want generation and verification in one recorded workflow.
RESOLVE("Build a small browser app that shows a keyboard-controlled square on a canvas","Must include complete runnable files, no build step, and JavaScript must pass syntax checks")>GEN(file_manifest,3)>JUDGE>UNFOLD_JUDGE(file_manifest,file_implementation,2)>MATERIALIZE(runs/canvas-square)>VERIFY("find . -name '*.js' -print0 | xargs -0 -n1 node --check")
Run trusted verification workflows with:
rick 'RESOLVE("Build a small browser app that shows a keyboard-controlled square on a canvas","Must include complete runnable files, no build step, and JavaScript must pass syntax checks")>GEN(file_manifest,3)>JUDGE>UNFOLD_JUDGE(file_manifest,file_implementation,2)>MATERIALIZE(runs/canvas-square)>VERIFY("find . -name '\''*.js'\'' -print0 | xargs -0 -n1 node --check")' --mode LOG_STEP --run-dir runs/canvas-square-run --max-calls 20 --allow-verifyRick reads environment variables from your shell and loads .env from the current directory. This keeps configuration local to the project where you run the tool.
OPENROUTER_API_KEY=
OPENROUTER_MODEL=google/gemini-2.0-flash-001
OPENROUTER_MODEL_LOW=openrouter/free
OPENROUTER_MODEL_MEDIUM=google/gemini-2.0-flash-001
OPENROUTER_MODEL_HIGH=google/gemini-3.1-flash-lite-preview
RICK_MAX_CALLS=60Security defaults block risky operations unless explicitly allowed:
CONTEXToutside the current directory- custom
OPENROUTER_BASE_URL VERIFYshell commandsMATERIALIZEoutsideruns/- materializing dotfiles or overwriting existing files
git clone https://github.com/para1992/rick.git
cd rick
python3 -m pip install -r requirements.lock -e .
python3 -m pytest
python3 -m rick_cli --help
python3 -m pip wheel --no-deps . -w /tmp/rick-release-wheelSee RELEASE.md for the release checklist.
Rick is early 0.1.1 alpha software. The current focus is functional quality, workflow reliability, and useful regression packs.
MIT.