Start at Step 0. Move down the page. Stop at every checkpoint.
Build a writing system.
One step at a time.
Use this page as a process: make a small source set, turn it into a System Card, test that Card on new work, then keep one owner-approved correction.
The live session, start to finish.
Forty-three minutes of this same page being worked through, including the parts that did not go to plan. Read the steps below, or watch them happen first.
What you are building
A writing system with rules and a definition of done, so you stop re-explaining your voice every session.
Pull the rules out of what you already wrote
Step 2 in practice: hand the AI emails and posts you have already written and let it name the rules you follow without noticing.
Inside a real System Card
A walk through the gates a draft has to pass: attribution, sources, fact checks, channel rules, voice match, and a person making the final call.
Fix the system, not the draft
The turn the whole session rests on. Feedback goes into the skill rather than into one output, so the next run starts further along.
Evaluations, and two ways to build them
What an eval is, a prompt-based version and a code-based one, and the open repositories you can copy them from.
Let your AI hold the process, not the decisions.
Paste this once if you want ChatGPT, Claude, Cowork, Codex, or Claude Code to take you through the page one checkpoint at a time. It can keep the log. It cannot fill a gap, approve a rule, or move on without you.
Before the page is live: use this prompt only in the rehearsal environment. The copied prompt is ready for the public URL after release.
Read the facilitator prompt
Act as my live workshop facilitator for: https://www.iwoszapar.com/webinar/write-in-your-voice Run exactly one numbered step at a time. Start at Step 0. For every step: 1. State the outcome in one sentence. 2. Explain why this step matters in plain language. 3. Ask me only for the information or decision needed for this step. 4. Give one safe example or default, but do not use it unless I explicitly say "use the example." 5. Stop and wait for my reply. Do not move on until I explicitly say "continue" or make the required decision. 6. Keep a short visible progress log: Done / Waiting for me / Next. Do not invent sources, personal context, opinions, or a fictional demo unless I explicitly say "demo mode." Do not publish, send, schedule, connect tools, or package a skill. At Step 1, ask me to choose one track: A) Iwo-attributed content, only if I can access the canonical local iwo-voice skill. B) Another person's or team's writing system. If I choose A, ask for the final channel. Before creating Iwo-attributed writing: - read the canonical iwo-voice skill; - ask me for the Content Origin Brief; - capture two verified content units, one author move, and the reader job; - choose a Register Card with me; - use only 2 to 4 matched approved Iwo examples; - show sentence provenance for substantive drafts; - run the appropriate phrase validator before presenting a final draft. After the System Card is confirmed for Track A, use the existing channel owner: - LinkedIn: write-linkedin - newsletter: write-newsletter - blog: publish-blog - email: draft-email If I choose B, do not imitate Iwo or claim access to his private system. Use only the participant's approved sources and the public workshop skills. Begin now with Step 0 only.
Bring three things.
- 01
One recurring job
For example: a proposal, a newsletter, or a project update.
- 02
Two or three sources
Only examples you trust and are allowed to place in your chosen AI.
- 03
One new brief
A comparable piece you can draft after the Card is confirmed.
Set up a private place for the system.
This is the first half of setup. Pick one private place for the sources and confirmed System Card. Use the option you already have. It does not change the process that follows.
Use a private Project or one fresh chat.
- 1. Name it
Writing System Lab. - 2. Add safe sources and the confirmed Card when the guide asks for them.
- 3. Paste each skill instruction into the same chat. Projects hold context, but do not package repository skills.
Use one local folder.
- 1. Name it
writing-system-lab. - 2. Keep
sources/andSYSTEM_CARD.mdinside it. - 3. Ask the agent to read the linked
SKILL.mdfile for each step. Do not install or configure a coding host during the workshop.
You have one private place to return to, and you know where the confirmed System Card will be saved.
Finish setup with one small source set.
Keep this deliberately small. You are making a system for one repeated writing job, not trying to extract your entire personality from everything you have ever written.
If the writing will be attributed to Iwo and you have the canonical local iwo-voice skill, name the final channel before you collect sources.
Build from the participant’s own approved examples. Do not imitate Iwo or claim access to a private voice system.
- 1. Name the recurring job.
- 2. Write the reader and the decision they need to make.
- 3. Collect two or three approved examples for that job.
- 4. Remove client contracts, credentials, regulated data, and anything you should not keep in this AI.
Use the fictional workshop pack.
It includes a proposal, a short newsletter, and a process memo. They are safe examples for the live build.
Open starter source packYou can name the job, reader, decision, and exact sources your AI is allowed to use.
Voice is evidence, then a decision, then language.
The useful part of iwo-voice is not a list of phrases. It stops examples from supplying the thought. A real situation supplies the thought. Examples only calibrate how that thought is said.
- 01
Reality
Content Origin Brief: origin, relationship, reader job, permitted facts, earned detail, and openness.
- 02
Thought ownership
Two verified content units, one author move, and the reader’s job exist before an example can shape the prose.
- 03
Register
Choose stance, energy, density, and ending. A public take and a careful email do not need the same voice setting.
- 04
Matched examples
Use only two to four examples matched by channel, relationship, job, and emotional temperature. They calibrate pace and directness, never the claim.
- 05
Provenance
For substantive drafts, each non-trivial sentence earns a reason: source, judgment, or reader need.
- 06
Hygiene last
The phrase validator catches known patterns. A clean scan means ready for judgment, not proof that the writing sounds like Iwo.
The point: do not add a marker to make a draft feel human. If the source, judgment, or situation is missing, stop and get that first.
Extract a Draft Contract from the sources.
The first skill does not write the final piece. It reads your sources and proposes the rules the future draft will have to meet.
writing-draft-contract
ChatGPT or Claude: open Raw and paste the instruction into your chat. Cowork, Codex, or Claude Code: point the agent at this public file, or paste it. The skill does not add an integration or save itself.
- Your chosen writing job and trigger.
- The reader and the decision the piece should support.
- Two or three approved sources.
- A proposed ten-field Draft Contract.
- Standards with source evidence, a pass test, and a failure test.
- Visible unknowns and a request for your confirmation.
The AI has asked you to keep, change, or remove its proposed standards. It has not drafted the final piece.
Confirm the Draft Contract, rule by rule.
This is your decision point. The AI can propose a rule. You decide whether it belongs in your system.
- 1. Keep the rules that are true and useful.
- 2. Change rules that are too vague to test.
- 3. Remove rules you do not want to maintain.
- 4. Mark anything unresolved as
not decided yet.
Do not fill gaps with plausible guesses.
If time is short, confirm fields 1 to 3, three standards across fields 4 to 6, and fields 7 to 9. Keep field 10 and unresolved preferences visible as not decided yet.
Every retained standard has an explicit keep, change, or remove decision. The Card is confirmed, but not yet placed in its permanent home.
Save the confirmed Card where future work can find it.
Put the same confirmed Card beside the same approved sources. This is the context your next draft will use. Do not write the new piece in this step.
Add the Card as a Project file and Project instruction. Without Projects, keep it in the chat and copy it into a plain document before leaving.
Add sources to Project knowledge and place the Card in instructions or an attached file. Without Projects, paste the Card into the fresh-work prompt.
Keep sources/ and SYSTEM_CARD.md together in the local folder. This is storage for the system, not a packaged skill.
copy out before leaving. Do not pretend the context will persist.The confirmed System Card and its approved sources are together in one place, or you have a clearly labelled copy-out fallback.
Test the Card on a fresh brief.
Now use the system on a new piece of work, beyond the source set. The workshop tests whether the Card can guide a new draft.
writing-blank-brief
ChatGPT or Claude: open Raw and paste the instruction into your chat. Cowork, Codex, or Claude Code: point the agent at this public file, or paste it. The skill does not add an integration or save itself.
Name the requested output, reader, decision, facts that are known, and what remains missing. Keep it comparable to the job your Card covers.
The skill returns the draft and an Output-to-Card map. Keep both. The map is the drafter’s traceability record; Step 6 is the independent check.
You have a fresh draft and an Output-to-Card map. The Card itself has not been silently changed.
Check the draft before you rewrite it.
Do not ask for “make it better.” Ask where the draft meets or breaks the rules you already approved.
writing-quality-check
ChatGPT or Claude: open Raw and paste the instruction into your chat. Cowork, Codex, or Claude Code: point the agent at this public file, or paste it. The skill does not add an integration or save itself.
- 1. What already fits the System Card.
- 2. The specific sentence that breaks a stated rule.
- 3. Unsupported claims, assumptions, or missing inputs.
- 4. One correction that will improve the next comparable run most.
You have chosen one specific correction. Do not rewrite the draft in this step.
A rule is useful only if a new draft can break it.
An evaluation is a small, deliberate test. It makes a future AI show whether it followed the rule instead of merely saying that it did. Today you build the raw material for an evaluation. Later, those same cases can become an evals fixture in a packaged skill.
Name one observable behavior.
Example: an outcome without source support stays an open question, not a promise.
Keep one failure and one pass condition.
The v1 unsupported sentence is the failure. The corrected, source-bounded sentence is a candidate for a future recheck.
Run it on comparable work.
The Output-to-Card map records the drafter’s traceability. The separate quality check decides whether the rule actually held.
- One rule with a pass condition.
- One exact failure sentence or missing input.
- One comparable next brief.
One corrected draft does not prove a stable system. Keep it recheck-pending, then run the same rule on comparable new work before you package anything.
Save one correction, with an approval gate.
A good correction becomes a rule only when you approve it. This last step stops the AI from quietly editing the system behind your back.
writing-save-correction
ChatGPT or Claude: open Raw and paste the instruction into your chat. Cowork, Codex, or Claude Code: point the agent at this public file, or paste it. The skill does not add an integration or save itself.
The skill shows the old rule, proposed new rule, pass test, failure test, and why this change is needed. Stop and review it.
Only after you approve the named change, update that Card field and revise the draft from the same brief. Keep the change map, then mark the revision recheck-pending.
You have a saved, owner-approved correction and a short map showing what changed and what remains uncertain.
Name the next manual proof.
One output is promising. The system earns trust when it handles comparable new work without you restating the rules.
- 01
Keep the Card and sources together.
- 02
Open clean work, or return for the next manual run.
- 03
Give it a comparable new brief with no restated rules.
- 04
Ask for the Output-to-Card map.
- 05
Keep, change, or stop based on the result.
Field 10 names one 15-minute manual test within seven days, with the exact next brief and run date.
See the advanced route. Do not start it today.
A repository skill packages a proven system. It is not a substitute for the Card, the manual runs, or your review.
- Today
Confirmed Card
One source-backed, user-approved system.
- Next
Three manual runs
Keep the corrections, failures, and source boundaries.
- Then
Skill Spec
Describe inputs, workflow, checks, exceptions, and an evaluation case.
- Package
skill-creator
Use Codex or Claude Code to test a proposed
SKILL.md.
Add the Card to the existing owner. Do not create a second publisher.
- 1. Keep the confirmed Card as a writing-standard input.
- 2. Let the current brief and research support the facts in the new piece.
- 3. Run the usual channel quality check and exact owner approval before anything is sent or published.
In my own system, that means /content selects one channel owner, then the normal voice, truth, evaluation, approval, and publishing rails stay in control. The Card can guide a draft. It never authorizes a claim, a send, or a publish.
ChatGPT and Claude Projects are complete for today: they can hold and run the Card. Do not install, configure, or expose a private repository during this workshop.
Open skill-creatorClose with one honest next run.
You leave with a small writing system, not a promise of automatic writing or a finished skill.
- 1. Keep the Card, sources, and correction together.
- 2. Put the exact next manual brief and its date in field 10.
- 3. Ask: What did the Card still fail to decide?
This guide is the working handoff.
Use the public skill links above now. A downloadable workshop pack should be added here only after its delivery route is live.
You have named a real next manual run. Package the system only after it has earned that step.
Keep these four artefacts together.
- 01
Confirmed System Card
The source boundaries and standards your future drafts must follow.
- 02
Fresh-work draft and Output-to-Card map
Evidence of how the system behaved on new work.
- 03
Quality report
The exact passes, breaks, and missing evidence.
- 04
One approved correction
A rule you can test again on the next comparable brief.
After three comparable runs, use skill-creator to adapt the public skills for your own work. If you already run a content system, pass the Card into its existing channel owner. Keep its evidence, evaluation, approval, and publishing gates intact.