Reduce AI-like patterns without losing what you meant.
Inception gives equal weight to two goals: reducing perceived AI-like convergence and preserving author intent. It is for drafts that feel formulaic, over-explained, repetitive, too neatly resolved, or more certain than their evidence.
It looks beyond word choice and tone, clusters co-occurring model-default features into upstream content decisions, and explains how those clusters create the effect. It then compares each decision with the author's purpose and constraints before proposing structural alternatives.
It does not rewrite the whole draft as soon as it finds a possible issue.
Inception first tells you:
- what it noticed;
- why the choice may be worth examining;
- how it relates to your original intent;
- how you could change it;
- what the change might cost.
You decide which parts actually need to change.
The first release supports:
- fiction and narrative writing;
- presentation text, including deck argument, slide function, claims, and evidence;
- document text, including reports, proposals, memos, specifications, policies, SOPs, and research summaries;
- organic brand and personal social copy, including captions, posts, threads, and post series.
Inception works with OpenAI Codex and Anthropic Claude Code.
Use $inception to review this draft for AI-like feature clusters
and intent conflicts.
Show me only the three most important decisions and wait before revising.
Inception defaults to a Quick Review: a focused set of evidence-backed decisions in natural language, usually one to three but with no hard maximum. It does not force every draft through a formal audit or hide a material issue to preserve the usual count.
For example:
Before
"The migration did more than improve our process. It transformed how
we think about collaboration. Here are three lessons every team can use..."
Inception notices
The opening claim, three lessons, and closing takeaway all resolve the
incident toward the same universal interpretation, while the draft
contains evidence from only one migration.
Human decision
Keep the incident. Remove the universal transformation claim and merge
the repeated lesson and takeaway.
After
"The migration took two days longer than planned because our rollback
assumption had never been tested under partial failure..."
The revision changes the content decision—not merely the synonyms—and happens only after the human accepts it.
Many humanizers mainly operate on surface-level writing, such as:
- swapping words;
- changing sentence length;
- removing clichés;
- making the tone more conversational;
- making sentence patterns less regular.
These techniques can sometimes help.
But many features that make writing feel “AI-like” appear further upstream.
Inception looks beyond:
How is this sentence written?
It also asks:
Why was this choice made here?
Does this information need to be here?
Does this conclusion need to be this complete?
Does this point need to be made again?
Is the claim as certain as the evidence supports?
Every conflict in a story is resolved through the protagonist's growth, understanding, and reconciliation, and the ending explicitly says what the protagonist “learned.”
Inception may point out:
- whether the ending is too neat;
- whether the theme is explained too completely;
- whether every conflict is being closed through the same resolution mechanism.
It does not decide that these patterns are wrong just because they are common in model outputs.
If the work is meant to be a fable, a healing story, or a story that needs clear closure, keeping them may be the right choice.
A presentation repeats the same conclusion across several slides and becomes increasingly certain, even though the evidence supports a more cautious claim.
Inception may point out:
- whether multiple slides are performing the same argument function;
- whether claim strength exceeds the evidence;
- whether the conclusion is stronger than the original intent.
A proposal explains the same rationale again and again in the Executive Summary, Background, Recommendation, and Conclusion.
Inception does not just rewrite the four passages with different sentence patterns.
It first asks:
Does this reason really need to be represented four times?
In other words, it checks the representation decision, not just the wording.
A post naturally falls into this familiar sequence:
hook
→ problem
→ personal experience
→ three takeaways
→ positive takeaway
If the author's intent was a cooler, more observational voice without a clear conclusion, Inception may flag this familiar structure as a content choice worth reconsidering.
It does not force the writing to become conversational or deliberately messy just because it does not “sound human” enough.
Most humanizers mainly operate on:
word choice
sentence structure
tone
register
Inception works further upstream:
content selection
argument structure
narrative resolution
claim strength
repetition
emphasis
explanation
Inception is not trying to make every piece of writing unusual, strange, or unpredictable.
Familiar choices are not automatically bad.
Unusual choices are not automatically better.
The real question is:
Is this choice serving the author's intent, or is it simply the easiest default for a model to choose?
For example:
- linear storytelling may be exactly what a children's story needs;
- a clear conclusion may be exactly what a policy document needs;
- repeating a key message in a presentation may be deliberate;
- a highly structured format may be exactly what an SOP needs.
Inception does not treat “common” as a synonym for “defective.”
It treats these patterns as decision candidates worth examining.
Most requests use Quick Review:
Draft → prioritized material decisions → you decide → revision → check
Quick Review infers a minimal intent snapshot from the request and draft, asks only when a missing answer could change the recommendation, and keeps the audit in the main conversation. It usually focuses on one to three decisions but has no hard finding cap.
Use Deep Audit when you ask for a deep, full, exhaustive, independent, or persisted review, or when a long or high-stakes artifact makes cross-section dependencies central:
Full Intent Contract
→ independent audit when available
→ Decision Ledger
→ you decide
→ authorized revision
→ Regression Audit
Finding count alone does not trigger Deep Audit. If more material candidates remain, Inception can continue Quick Review, run a focused follow-up, or recommend Deep Audit when dependencies, stakes, persistence, or independent review actually justify it.
Both depths first establish:
What is this piece of writing trying to achieve?
Only then does it examine the draft's content decisions.
If it finds important choices that may conflict with the intent, it returns them to you first.
You can:
- accept them;
- modify the recommendation;
- reject them;
- defer them.
Only decisions that you accept or modify can enter revision. Quick Review changes the amount of ceremony, not this human-decision gate.
After the revision, Inception checks again:
- whether the approved change was applied correctly;
- whether the change went beyond the authorized scope;
- whether it introduced a new problem or trade-off.
StoryScope (📄 arXiv paper) observed in its research on AI fiction that AI-generated stories may be more likely to use neat, single-path plots and explain their themes too explicitly.
Inception uses this kind of population-level observation as a reference for proposing audit hypotheses. It is not a writing rule for all content, and it does not mean that presentation text, document text, or social copy has the same level of research evidence.
This skill is not an AI detector or a cosmetic humanizer. Its purpose is to identify content decisions that may be influenced by default patterns, compare them with the author's original intent, and return decisions with real material impact to human judgment.
Suppose you are preparing a presentation for an engineering team.
Your intent is:
Audience:
Engineering team
Goal:
Help the team understand the limitations of the current approach
Must preserve:
Technical accuracy
Known uncertainty
Existing evidence
Must avoid:
Overstating the conclusion
The draft might contain this sentence:
This architecture solves the scalability problem and provides a reliable foundation for future growth.
But the current evidence has only been tested in a limited set of situations.
Inception might return:
Observed choice:
The conclusion describes the architecture as having fully solved the scalability problem.
Why it matters:
The original intent requires the limitations and uncertainty to remain visible.
Possible alternatives:
- Keep the original sentence;
- narrow the claim to what the current evidence supports;
- distinguish verified results from expected future benefits.
Recommendation:
Revise.
You still decide whether to accept the recommendation.
Inception does not silently change the original sentence and then tell you it has been “optimized.”
The project-local skill is stored at:
skills/inception/
├── SKILL.md
├── agents/openai.yaml
├── references/
│ ├── core-workflow.md
│ ├── document-text-adapter.md
│ ├── fiction-adapter.md
│ ├── presentation-text-adapter.md
│ ├── reviewer-protocol.md
│ └── social-copy-adapter.md
└── scripts/validate_ledger.py
SKILL.md loads the Core workflow and exactly one domain adapter. Deep Audit additionally loads the Reviewer Protocol; a deep fiction audit also loads the complete StoryScope feature registry.
More detailed domain-specific guidance remains in references/, so unrelated domain documents do not consume context for the current task.
The repository is the single source of truth for this skill. The Codex and Claude Code packages both load the same skills/inception/SKILL.md; neither package contains a separate copy of the skill implementation.
Integration metadata is kept at the repository root: Codex uses .codex-plugin/plugin.json and .agents/plugins/marketplace.json, while Claude Code uses .claude-plugin/plugin.json and .claude-plugin/marketplace.json.
Add the repository marketplace, then install the plugin:
codex plugin marketplace add arthur422tp/Inception
codex plugin add inception@inceptionAfter installation, invoke it explicitly with $inception, or ask Codex to audit or revise fiction, presentation text, document text, or organic brand and personal social copy against your intent.
Add the repository marketplace, then install the plugin:
claude plugin marketplace add arthur422tp/Inception
claude plugin install inception@inceptionInside an interactive Claude Code session, you can use these equivalent commands:
/plugin marketplace add arthur422tp/Inception
/plugin install inception@inception
After installation, the skill is available under the plugin namespace /inception:inception. For local development, load the current repository directly with claude --plugin-dir ..
For local development or older Codex versions without plugin marketplace support, copy or link the skill folder into ~/.codex/skills/:
cp -R skills/inception ~/.codex/skills/Or use a symbolic link while developing locally:
ln -s "$(pwd)/skills/inception" ~/.codex/skills/inceptionAfter installation, invoke it explicitly with:
$inception
Decision Ledgers use JSON, so workflow gates can be checked deterministically:
.venv/bin/python skills/inception/scripts/validate_ledger.py \
tests/fixtures/valid-ledger.jsonExpected output:
valid: tests/fixtures/valid-ledger.json
Invalid input exits with a non-zero status and reports the exact field path where the problem occurred.
The Python utilities use only the Python 3.11 standard library:
.venv/bin/python -m unittest discover -s tests -p 'test_*.py' -vYou can also use Codex's skill-creator utility to validate the skill folder itself:
python3 /Users/arthuryu/.codex/skills/.system/skill-creator/scripts/quick_validate.py \
skills/inceptionThis project is not intended for:
- AI-source detection or authorship scoring;
- word blacklists or other cosmetic humanization techniques;
- automatic rewriting before human disposition;
- autonomous reviewer-to-writer revision loops;
- treating StoryScope's population-level tendencies as universal writing rules for all text;
- visual presentation design;
- document layout, file-format operations, tracked changes, or comments;
- paid ads, targeting, bidding, social scheduling, or algorithm speculation.
MIT