Workflow Capture

Workflow Capture

Maintain an append-only journal that survives compaction and /clear, capturing the signal (decisions, sequencing, corrections, dead ends) that should become a skill. Do not log keystrokes; the transcript already does that at the wrong altitude. Capture why, not just what.

At session start

1. Determine the active workflow slug (kebab-case, e.g. figma-site-redesign).
2. Look for _worklogs/YYYY-MM-DD-/WORKLOG.md.

• If it exists, read it fully before doing anything. It is your memory.
• If not, create the directory and a new WORKLOG.md with a one-line header naming the workflow and date.

Tag schema

Every entry carries exactly one tag:

• DECISION: what was done and why -> a procedure step
• SEQUENCE: X must precede Y -> ordered steps or preconditions
• CORRECTION: user overrode you -> a hard rule
• DEADEND: tried X, failed because Y -> a "do not do this" warning
• CONVENTION: naming, structure, params, defaults -> reference material
• DISCOVERY: tool quirk, API limit, gotcha -> environment notes
• CHECKPOINT: state snapshot before clear or compaction -> resume anchor

Entry format

Append, never rewrite history.

## [YYYY-MM-DD HH:MM] TAG
One-line summary.
Optional result and fix or lesson.
Tags: comma, separated, keywords

When to append

Append an entry the moment one of these happens:

• The user corrects or overrides something (CORRECTION).
• An approach fails and you switch (DEADEND).
• A non-obvious ordering constraint appears (SEQUENCE).
• You set a reusable convention (CONVENTION).
• A tool behaves unexpectedly (DISCOVERY).

Keep entries short. The lesson matters, not the prose.

Checkpoint and resume

Before any compaction or /clear, write a CHECKPOINT entry: current state, open threads, next intended action. On resume in a fresh window, read the worklog and continue from the last CHECKPOINT plus accumulated tags.

Why this matters

The context window is volatile and mostly noise. A durable tagged journal is the only thing that lets a workflow survive a context reset and later become a skill that makes the next run cheaper and more correct.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Mfrostbutter
• Source: Mfrostbutter/skill-forge
• License: MIT

Install and usage instructions live in the source repository linked above.