FREESKILL v1.0UPDATED 2026-08

Writing with AI

Generation is not approval. Writing with AI turns the boundary between machine assistance and human judgment into a six-stage workflow with evidence checks, voice review, and a named approver.

A better prompt does not solve an undefined document job. Before generation, the workflow names the reader, requested outcome, source boundary, format, voice basis, and the decisions a person must own. Each stage then carries a human check and exit evidence, so a draft cannot become approved by momentum.

The second artifact is an AI Tell Audit. It treats generic framing, mechanical symmetry, unsupported certainty, and flattened voice as repairable text patterns, never as proof of authorship. Truth comes before style. Claims are checked against supplied sources, voice is checked only against a supplied basis, and the final verdict stays bounded to readiness for human review.

AI can generate options. It cannot own the purpose, evidence, judgment, or final approval.

SIGNALS YOU NEED THIS

A team generates drafts before agreeing on the reader, outcome, or allowed evidence.

A fluent factual claim reaches review without anyone being able to name its source.

Someone wants a style pattern converted into an AI-authorship accusation or confidence score.

REPLACES

One-shot prompting, untracked fact checking, generic make it sound human requests, and generation treated as approval.

PAIRS WITH

Business Writing, Changelog, and Done.

0 of 4 steps complete

STEP 01Define the document job and the evidence boundary

Name the reader, requested outcome, format, length, allowed sources, and what may not enter an AI system. Secrets, personal data, regulated material, and confidential content require approved handling or removal. A prompt cannot compensate for a document with no purpose or for evidence the workflow is not allowed to use.

DOCUMENT JOB
  Reader: department leads
  Outcome: surface decisions, risks, and owner changes
  Format: one-page weekly update

EVIDENCE BOUNDARY
  Allowed: supplied project notes and two approved voice examples
  Not allowed: confidential material outside those notes
  Missing claim support: completion date and measured result

STEP 02Reserve human decisions and build the six-stage loop

Identify the claims, tradeoffs, judgments, and approvals a person must supply. Then define Frame, Generate, Refine, Validate, Polish, and Approve. Each stage states what the human supplies, what AI may do, what the human checks, and the evidence required to exit. The Approve stage gives AI no approval authority.

FRAME      exits with an approved brief
GENERATE   exits with a candidate set inside the brief
REFINE     exits with a working draft whose meaning is checked
VALIDATE   exits with a claim inventory
POLISH     exits with a review draft checked against the voice basis
APPROVE    exits only with the named human's approval and date

No earlier stage inherits the authority of the final one.

STEP 03Verify claims, then audit observable patterns

List every factual assertion, figure, date, name, and quotation with its supplied source and status. Remove or label unsupported content before polishing it. Then locate generic framing, repeated setup, mechanical symmetry, flattened voice, decorative transitions, inflated stakes, unclear ownership, and term drift. Each finding needs a location, consequence, evidence, and repair.

TRUTH FIRST
  Claim: the project will improve every team's outcomes next month
  Supplied support: missing
  Action: remove or narrow after evidence is supplied

PATTERN FINDING
  Location: conclusion
  Pattern: universal outcome claim plus repeated generic stakes
  Reader consequence: confidence without a checkable basis
  Repair: state the supported decision or leave Evidence needed

STEP 04Restore voice from evidence and end with a human verdict

Compare the draft only with voice examples or explicit rules the user supplied. If none exist, write No voice basis supplied. Prioritize truth and changed meaning before style, assign no more than three repairs, and choose Ready for human review, Revise, or Rebuild from the brief. Never convert pattern counts into an authorship probability.

VOICE BASIS
  Two supplied leadership updates

PRIORITY REPAIRS
  1. Remove the unsupported result. Owner: Owner needed.
  2. Replace the generic conclusion with the requested decision.
  3. Restore the supplied update owner's direct terms.

VERDICT
  Revise

This verdict concerns readiness. It does not identify the writer.

Download the folder into your agent's skills directory. It uses only the material you provide and makes no remote calls.

skills/writing-with-ai

writing-with-ai/
├── SKILL.md                         ← workflow design and audit modes
├── meta.yaml                        ← version, sources, test prompts, changelog
├── LICENSE.md                       ← MIT
├── assets/
│   ├── drafting-workflow-spec.md    ← the six-stage human and AI loop
│   └── ai-tell-audit.md             ← truth, voice, and pattern findings
└── references/
    └── review-patterns.md           ← audit order and repair rules

INVOKE IT LIKE THIS

"Set up an AI drafting workflow that still sounds like me."

Download skill v1.0 View source files

WORKS WITHClaude Code