Evolution from RPI to CRISPY: Multi-Stage Workflow for Production Coding Agents

humanlayer · Tech · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: production coding-agent workflow design

HumanLayer needed to make its Research-Plan-Implement methodology work across real engineering teams after a giant planning prompt proved inconsistent, hard to review, and too dependent on expert prompting.

Constraints:

• The original planning prompt had about 85 or more instructions.
• Frontier LLMs may only follow roughly 150-200 instructions consistently.
• Teams still need code review and human-agent alignment to preserve quality.
• Thousand-line plans were often as long as the code and could hide surprises.
• Methods that work for expert engineers did not automatically scale to enterprise teams.

Architecture

HumanLayer evolved RPI into a staged CRISPY-like workflow: questions, research, design discussion, structure outline, plan, work tree, implement, and PR. Instead of one monolithic prompt, each stage has fewer than about 40 instructions and produces intermediate artifacts. Design discussions and structure outlines become shorter human-review points, while final code review remains mandatory rather than assuming autonomous generation is enough.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  question-and-research-stages["Question and research stages"] --> design-discussion["Design discussion"]
  design-discussion --> structure-outline["Structure outline"]
  structure-outline --> tactical-plan["Tactical plan"]
  tactical-plan --> implementation-workspace["Implementation workspace"]
  implementation-workspace --> pull-request-review["Pull request review"]

Components

Data flow

The workflow starts by clarifying questions and researching the codebase, then opens a design discussion, captures resolved decisions and open questions, creates a structure outline, writes a tactical plan, prepares the work tree, implements the change, and sends a PR for review. Each stage uses a focused prompt and a fresh context window rather than one giant prompt controlling the whole process.

Why this architecture

HumanLayer split the workflow because prompts are a poor control-flow mechanism once instruction count grows. Focused stages reduce instruction burden, make human review documents shorter, and preserve quality by aligning on design and structure before code generation and mandatory review.

Models used: Claude Code · RPI prompts · CRISPY workflow · Open source prompts · Coding agents

Scale

• About 10,000 people grabbed the open-source prompts.
  “probably 10,000 people who have gone to our open source and grabbed our prompts”transcript, stated
• HumanLayer worked with thousands of engineers up to Fortune 500 companies.
  “basically worked with thousands of engineers from tiny startups all the way up to Fortune 500s.”transcript, stated
• The presenter cites roughly 150-200 instructions as the consistent-following range for frontier LLMs.
  “frontier LLMs could only follow about 150 to 200 instructions with like good consistency.”transcript, stated
• Structure outlines were about two pages versus roughly eight-page plans in one example.
  “Plan eight pages, structure outline two-ish pages, much shorter.”transcript, stated

Alternatives considered (roads not taken)

Keep one giant RPI planning prompt.

Rejected because: The prompt had too many instructions and skipped important alignment steps for many users.

Skip human review of intermediate artifacts.

Rejected because: Plans and code could diverge or surprise reviewers without design and structure alignment.

Fully autonomous code generation without code review.

Rejected because: The speaker says quality still requires alignment, review, and rework.

Patterns

prompt-chaining (coding-agent workflow)

RPI is decomposed into smaller focused prompt stages rather than one mega-prompt.

human-in-the-loop-gate (engineering review)

Design discussions, structure outlines, and pull requests create review points before and after implementation.

context-compaction (coding-agent sessions)

The workflow creates artifacts that let sessions resume without relying only on automatic compaction quality.

Tradeoffs

Use multiple short artifacts.

Chose Design discussion and structure outline before plan over Reviewing only a giant final plan.

Rationale: Shorter artifacts give humans more leverage and earlier chances to redirect.

Keep code review mandatory.

Chose PR review after implementation over Trusting autonomous generation once the plan is approved.

Rationale: Even with better planning, teams still need human code ownership and rework loops.

Gotchas (what bit them)

Instruction overload caused skipped workflow steps.

Root cause: The model had to follow the big RPI prompt plus system prompt, tools, Claude.md, and MCP instructions.

Resolution: Split the workflow into prompts under roughly 40 instructions each.

Thousand-line plans were too costly to review.

Root cause: The plan could approach the size of the resulting code and still contain surprises.

Resolution: Use shorter design discussions and structure outlines as higher-leverage review artifacts.

Models kept writing horizontal plans.

Root cause: Prompting alone did not reliably make agents produce vertically sliced, testable implementation plans.

Resolution: Add a structure-outline stage that breaks the work into reviewable blocks before the tactical plan.

Evaluation

The workflow was informed by workshops with engineers, prompt behavior observation, instruction-following limits, and attempts to prompt/eval planning behavior such as avoiding horizontal plans.

Metrics: Instruction adherence · Review document length · Human-agent alignment · Code review rework · Plan quality · Team adoption

Outcomes

• HumanLayer replaced the monolithic RPI planning prompt with staged questions, research, design, structure, plan, work tree, implement, and PR steps.
• Each split prompt dropped below roughly 40 instructions.
• Shorter design and structure artifacts improved human-agent and team alignment before code review.

The talk reports 10,000 prompt downloads and workshops with thousands of engineers from startups through Fortune 500 companies.