llearn

Independent applied AI practice

Working model

Method

Day 95 of iteration since : a project keeps durable truth, then runs bounded work through an explicit workflow.

Foundations hold cross-loop truth. A workflow is reusable instruction. A loop is one bounded execution with an append-only state trail.

Project truth

A project owns the facts, decisions, and language that must survive the current session. Those foundations stay concise and cross-loop; scratch is a capture lane, not a source of truth.

Foundations
The brief, guardrails, success evidence, and project-local language. They explain what the project is, what must hold, and what words mean here.
Startup index
Exact, task-relevant targets for a restart: foundations, the selected workflow, and the active loop’s charter and state trail. It prevents work from relying on remembered conversation.
Language
Small definitions for terms that change action, output, safety, privacy, or boundaries. Name important distinctions before agents drift around them.
Project index targets The project index points to three equal target classes: foundations, the selected workflow, and the active loop. Project index exact targets for this job Project truth Foundations brief · guardrails success · language Named process Selected workflow exact path version or revision Current work Active loop loop.md state.jsonl Project index targets The project index points to three equal target classes: foundations, the selected workflow, and the active loop. Project index exact targets for this job Project truth Foundations brief · guardrails success · language Named process Selected workflow exact path version or revision Current work Active loop loop.md · state.jsonl
At startup, the index points to project truth, the named workflow, and the active loop. Scratch is not a startup target.

Guide topics: Context Engineering and RAG.

Bounded loops

A loop is one execution of a workflow, not a permanent workspace. The workflow is named by path and fixed by a version or revision so a later session can understand what it was following.

One bounded loop A single left-to-right sequence with three equal stages: Open, Work, and Close. Open 1 write loop.md state: active Work 2 use working/ if useful append meaningful state Close 3 state: closed or cancelled write final/summary.md One bounded loop A single top-to-bottom sequence with three equal stages: Open, Work, and Close. 1 Open write loop.md state: active 2 Work use working/ if useful append meaningful state 3 Close state: closed or cancelled write final/summary.md
The workflow governs one bounded execution. The state trail records only meaningful transitions from open to close.

  1. Open

    Write loop.md: the goal, inputs, loop-specific guardrails, exit criteria, permitted deviations, and workflow reference. Append an opening state.jsonl row with status active.

  2. Work

    Create mutable working material only when an artifact is useful. Append state only for meaningful transitions: a decision, validation, blocker, handoff, material reframing, or status change—not routine reads or retries.

  3. Close

    Append a closed or cancelled transition, write final/summary.md, and route durable output deliberately: truth to foundations, procedures to workflows, support to resources, explanation to docs, and provenance to receipts.

Where finished work goes
If the result is File it in
Project truth foundations/
Repeatable procedure workflows/
Reusable support resources/
Explanation or reference docs/
Decision or provenance receipts/
At closeout, file each lasting result in one project folder. Scratch remains temporary.

When work stalls, reduce motion rather than add ceremony: write a synthesis, inspect assumptions, record a blocker, reframe, delegate, or shrink the step.

Guide topic: Loop Engineering.

State

state.jsonl is append-only. Its latest valid row is authoritative for continuation and close state. Each meaningful row records a timestamp, loop ID, status, factual summary, references, and next action; corrections are new rows, never rewrites of history.

Earlier events.jsonl loops remain valid historical evidence. New loops use the lean loop.md and state.jsonl model.

Keep truth durable. Select the workflow. Run the bounded loop. Record meaningful transitions. Leave the next session a clear place to begin.

Explore the Practical AI Guide