Claims first, prose last, numerals only once.

Stage by stage

Eight stages, each one resumable and each one correctable.

Progress is tracked per stage, so a draft can be left and picked up. The percentages below are the pipeline's own — they are weights on the work, not a guess at how long it takes.

1. 1

   Invention structure

   IDEA_ENTRY · 10%

   The disclosure is turned into structure — what the invention is, what it operates on, what problem it solves. Everything downstream reads this, so it is the first and cheapest place to correct a misreading.

2. 2

   Preliminary claims

   PRELIMINARY_CLAIMS · 22%

   Claims are drafted early, deliberately before the prose. The claims decide what the specification has to support, so writing the description first means rewriting it later.

3. 3

   Prior art analysis

   RELATED_ART · 35%

   Prior art is brought in and reviewed against the preliminary claims — either handed over from a novelty search or entered directly. This is what the background section will later be built from.

4. 4

   Claim refinement

   CLAIM_REFINEMENT · 45%

   Now that the art is known, the claims are narrowed and layered — independent scope tested against the closest references, dependents added as fallback positions.

5. 5

   Component planner

   COMPONENT_PLANNER · 58%

   Every component of the invention is enumerated and assigned its reference numeral once, centrally. This is why numerals stay consistent between the figures and the description instead of drifting apart.

6. 6

   Figure planner

   FIGURE_PLANNER · 70%

   The figure set is planned as a set — which figures exist, in what order, of which kind — then each is generated, validated for filing readiness, and sequenced. Diagrams and hand sketches sit in the same sequence.

7. 7

   Jurisdiction setup

   COUNTRY_WISE_DRAFTING · 55%

   The target offices are chosen, and each one contributes its own section list, ordering, and word limits drawn from a superset of canonical sections. One invention, several jurisdiction shapes.

8. 8

   Draft sections

   ANNEXURE_DRAFT · 82%

   Sections are drafted one at a time, each receiving only the context its own definition asks for — and each checked against the sections around it rather than in isolation.

Context injection

Each section is told only what it needs to know.

A section is not a prompt with everything thrown at it. Every canonical section declares which context it requires, and the drafting call receives exactly that — which is both why the output stays on-topic and why a section can be redrafted without disturbing its neighbours.

Each section also carries its own base instruction, its constraints, whether it is required, and a recommended word range — so a jurisdiction can reorder or drop sections without the drafting logic being rewritten.

Figure planner

Four kinds of figure, because a method claim and an assembly are not the same drawing.

The figure set is planned before anything is drawn — which figures, in what order, of which kind. Each kind has its own schema, so a sequence diagram is validated as a sequence rather than as a generic box-and-arrow picture. Relationships inside a figure are typed rather than left as anonymous lines — primary, data input, technical output, control, configuration, validation, storage, optional — which is what lets a drawing be checked for sense instead of merely rendered.

Filing readiness

A figure that an examiner would object to is caught before you file it.

Every diagram is validated, not just rendered. The report is explicit about whether the figure is filing-ready, which claim-critical components are missing from it, and whether it has become too dense to be one sheet.

Your figures, your way

Generated, overridden, or imported — all in one sequence.

A managed figure regenerates when the components change. The moment you edit its source by hand, that edit is respected instead of being overwritten on the next run — and imported images and sketches take their place in the same numbered sequence.

Voice

Drafted in your style, not the model's.

Writing samples are stored per jurisdiction, per persona, and per section, because how you write claims has nothing to do with how you write a background. Five sections carry most of the voice, and their readiness is tracked separately so you know what is actually configured.

Review

Sections are checked against each other, not one at a time.

The interesting defects in a specification live between sections: a claim feature with no home in the description, a numeral in a figure that the prose never mentions. Cross-section rules run from a source section to a target section, and the fix is pointed at the section that has to change — not the one that revealed the problem.

One invention, several offices

The section list is the jurisdiction's, not ours.

Canonical sections sit in a superset with aliases, display order, and per-office word limits. A jurisdiction takes the sections it wants in the order it wants them — objects of the invention here, technical problem and solution there, industrial applicability where it is required — and the same drafted invention comes out shaped correctly for each.