WTF-P

The question

Language models can produce a plausible paragraph almost instantly. That is not the same as developing an argument, locating its evidence, controlling its scope, or noticing when one section quietly contradicts another.

WTF-P started from a recurring problem inside real research work: the blank page was only the visible symptom. Underneath it were missing decisions about the question, contribution, evidence inventory, narrative, venue, and ownership. Could an AI-assisted writing tool make those decisions explicit before asking for prose?

The workflow bet

The bet is to treat writing as a staged system. A project begins with a specification. The paper becomes an outline with an argument and word budget. A section receives its own plan and bounded context. Drafting is followed by a review that works backward from the section’s stated purpose and available evidence.

This is less magical than “write my paper,” and deliberately so. It gives the human researcher places to inspect and change the reasoning. It also makes failure more local: a weak paragraph can point back to a missing source, an incoherent section plan, or a project claim that was never resolved.

What shipped

WTF-P v0.5.0 is a public npm package and GitHub release. Its installer can place a common writing workflow into Claude Code, Gemini CLI, or OpenCode. The repository includes commands for project setup, planning, drafting, review, structure, and export; specialized agent roles; tests; a changelog; and a continuous-integration check at the released commit.

The package is the proof that this moved beyond a private prompt folder. Installation, updates, diagnostics, runtime differences, and user-owned configuration all became product questions. Solving them changed the work from a clever instruction into a tool another researcher could try and remove.

My part, and why I built it

I created WTF-P from the writing and mentoring workflows around GRC and maintain its architecture, installer, commands, and release path. The deliberately blunt name comes from deadline energy, but the system underneath it is patient: decide what a section must accomplish, gather the evidence it may use, write inside that boundary, and review the result.

This is also entrepreneurial practice at a small and useful scale. I saw a repeated operational problem, formed an opinion about how it should work, packaged the opinion, and exposed it to real use. The point is not to turn research judgment into a service. It is to make more of that judgment visible at the moment a team needs it.

The early boundary

WTF-P is early software below a 1.0 release. Its workflows do not verify that a scientific claim is true, make a weak source authoritative, or turn generated text into the author’s considered argument. Citation grounding and review steps can create better checkpoints; they cannot substitute for reading the sources or owning the final claim.

The released package also represents one opinionated workflow, not a universal model of scholarship. Disciplines, venues, teams, and individual writers organize evidence differently. The useful test is whether the structure helps a researcher see and improve the work, not whether every paper follows the same sequence.

What I am learning

• Context quality depends more on explicit boundaries than on putting the entire project into one prompt.
• Section isolation is useful only when the shared argument remains visible.
• Installers and diagnostics are part of an intellectual tool because they determine who can actually try the idea.
• The right role for an agent is often to maintain a demanding process, not to impersonate the author.