FlowPrint

• 

  logo

Inspiration：

Skills are appearing everywhere, but I kept coming back to a more basic question: what actually makes a good Skill? I also wondered whether the repeated process of working through a problem in a chat—trying something, correcting it, and trying again—could be distilled into the useful parts of a reusable Skill instead of being lost inside the conversation. That question became the starting point for FlowPrint.

My first real example came from a sticker-making workflow for Lao San, my family's pet. I wanted to see whether the decisions and corrections from that creative process could become something reusable without copying all of the personal details from the original conversation. Later, I tested FlowPrint on a family travel conversation and tried to extract a more general trip-planning Skill. The result was not as good as I hoped, which exposed problems in workflow scope and generated structure and sent me back into another round of revisions.

FlowPrint grew from those practical failures. It is my attempt to make reuse deliberate: preserve the useful method, keep one-time context separate, and let the user inspect what will be preserved before a Skill draft is compiled.

FlowPrint was built to make reuse deliberate. It separates model-based language understanding from deterministic enforcement, so a user can inspect what will be preserved before any Skill draft is compiled.

What it does：

FlowPrint starts after an AI-assisted task is complete. It discovers candidate reusable workflows and stops for user selection when one conversation contains multiple outputs. It then classifies the selected workflow into six layers:

Core Workflow Domain Knowledge Profile Run Parameters Failure Lessons Permission Boundaries After review and explicit confirmation, FlowPrint compiles a non-installed Skill draft. Deterministic gates validate schema state, workflow scope, confirmation records, dependency fingerprints, revision receipts, and permission status. A draft is never treated as authorization to install the Skill or perform external actions.

How we built it：

GPT-5.6 handles evidence interpretation, workflow-candidate discovery, decontextualization, and natural-language correction analysis. Python scripts own deterministic validation and compilation.

The compiler and validators enforce:

broad-directory evidence scope blocking; explicit workflow selection for multi-workflow conversations; schema and confirmation gates; fail-closed compilation with private staging; immutable base drafts during revision; dependency fingerprints and revision receipts; and separate authorization for compilation, installation, and external actions. The plugin is packaged for Codex and includes PowerShell support for Windows. The public repository contains fixtures, evidence records, platform documentation, and automated regression tests.

Challenges we ran into：

The hardest problem was not generating Markdown. It was deciding what information was genuinely reusable, what belonged only to one run, and what should never be carried into a future Skill.

The Lao San sticker workflow was a useful initial case because it included concrete user corrections. The family travel test was harder. One conversation contained both trip planning and poster production, and the first generated Skill mixed them together. Even after separating the workflows, the result could still look structurally correct while reading more like a summary than something ready to execute. I revised the pipeline to discover and select workflows before six-layer classification, then added clearer input, output, and quality contracts to generated drafts.

Cross-platform acceptance testing also exposed evidence-scope and version-cohort problems. FlowPrint now blocks recursive discovery from broad personal roots and requires evidence claims to remain tied to the selected test cohort.

I also tried not to follow every AI suggestion blindly. At each important stage, I asked colleagues with relevant professional experience to review the evidence and engineering judgment. I compared their views, decided what I agreed with, and only then moved to the next step. That made the process slower, but it helped me keep my own reasoning in the loop.

Accomplishments that we're proud of：

A clear separation between GPT-5.6 language understanding and deterministic Python gates. Fail-closed workflow-selection, confirmation, compilation, and revision paths. Separate permissions for draft compilation, Skill installation, and external actions. Real Windows and macOS acceptance evidence for the tested paths and versions. A held-out sticker workflow accepted after two user-guided correction cycles. A public repository with 63 structural and regression tests. The test count describes deterministic engineering coverage. It is not a claim of statistical classification accuracy or general generated-content quality.

What we learned：

I learned that a useful Skill is not simply a polished summary of a successful conversation. Reliable reuse requires explicit workflow scope, evidence provenance, user confirmation, artifact dependencies, and permission boundaries.

I also learned that safety gates and output usefulness have to evolve together. A carefully protected draft is still not useful if it lacks clear inputs, missing-information handling, an output contract, and checkable quality criteria. The disappointing travel result was valuable because it made that gap visible.

Most importantly, professional review is not a substitute for my own judgment. It works best as a way to challenge assumptions, compare perspectives, and make the next revision more deliberate.

What's next for FlowPrint：

• evaluate generated Skills on more independent real-world tasks, including office and data-analysis workflows;
• repeat the creative test with a different character to check that Lao San-specific details do not leak into new Skills;
• add more content-level tests for inputs, output contracts, and quality checks;
• improve the first-run experience when evidence-scope protection blocks a broad directory;
• expand cross-platform acceptance coverage without generalizing beyond tested versions; and
• prepare FlowPrint for the appropriate public plugin review channel when submission prerequisites are available.

Built With

• agent-skills
• ai-agents
• build
• codex
• developer-tools
• flowprint
• gpt-5.6
• open-source
• openai
• python
• workflow-automation