Train & learn

Training learns how you revise. Learn mode uses your real edits to do the same job. Both keep content corrections separate from voice evidence and ask before saving a lesson.

Run one training round

/vf-train explainer-en

The agent chooses a topic not used in the last five rounds and writes one or two paragraphs, normally 120 to 250 words. Rewrite them as you would write them today. You may answer OK to keep the passage, skip for a different topic or stop to end.

The passage should be competent and substantive. Ordinary rounds do not deliberately stuff the text with slop. A hard round is a separate, explicit option.

Your rewrite is saved exactly, including typos, before analysis. The agent checks for duplicates, shows up to six findings and proposes at most three lessons. Say yes, no or reword each one.

Vary the task

/vf-train explainer-en openings
/vf-train explainer-en closings
/vf-train explainer-en transitions
/vf-train explainer-en data
/vf-train explainer-en from essay-one.md

Own-text rounds use a paragraph you wrote; you rewrite it as you would today. They can help distinguish your current voice from an old habit or a coauthor's style. Focuses make one part more prominent without turning a single paragraph into a whole document.

From lesson to approved rule

A lesson starts as candidate. The evidence threshold is:

• Three or more unique rounds across at least two topics; or
• Two rounds on two topics, plus an analysis.

Learn rounds count like Train rounds. Duplicate rounds do not count again. One analysis can support the second path; additional analyses do not lower the round requirement.

Every five rounds, at a session's end or when you request status, the agent checks what is ready. Readiness is an invitation to review a rule. You must approve the complete contract: instruction, strength, kind, scope, exceptions and check. Understand those fields.

/vf-status explainer-en
/vf-rules explainer-en

A confirmation supports a rule. A contradiction is evidence to examine, rather than something to hide. Two or more training rounds against a rule should prompt a proposal to reword or drop it, with your approval.

Learn from two files

/vf-learn explainer-en draft.md final.md

The first file is the AI draft. The second is your final version. The agent compares changed paragraphs and a little context, saves a Learn round and follows the same findings and lesson process.

A changed deadline is content. A cut made only to meet a word limit is situational. Your coauthor's edits are not your voice. Tell the agent which changes are yours when attribution is unclear.

Learn from Git history

/vf-learn explainer-en essay.md

For a single file in a Git repository, baseline selection follows this order:

1. A vf-draft commit where Voice Forge wrote the draft, compared with the current file.
2. Your uncommitted changes, compared with the last commit.
3. The last two commits that changed that file.

The agent reports the two versions before learning. You can append an actual baseline commit to choose another comparison. If a file is outside Git, provide two files instead.

The helper flags other authors and already learned pairs. No changes or only one version means there is nothing to learn yet. Code, tables, data and references stay out of voice learning.

What gets stored

Rounds, lessons, topics and approved rules stay under the selected profile in your private voice home. A shared rule belongs in base.md only after you confirm it applies across your writing. One profile's rule never silently moves into another.

In a plain chat packet, the agent cannot save files. It prints round and lesson blocks for you to copy back to your private voice home. Full Train and Learn methods are available on this site.