Reframe

Methodology for designing in territory where no one yet knows the shape. When you're building "AI native CRM" or "agent-first IDE", the danger isn't that you fail outright — it's that you succeed at a traditional version wearing the new paradigm's clothes. New buttons, old skeleton. The user gets traditional capability with paradigm overhead: the worst of both worlds.

This skill enforces a sequence designed to prevent that. Distill what the domain actually does underneath its surface. Redraw the load- bearing skeleton using the new paradigm's primitives. Test whether the skeleton stands without traditional crutches. Then add the familiar surface as a projection of the new skeleton.

Commands

When invoked with an argument, dispatch to the corresponding file:

• /reframe init → Read and follow commands/init.md in this skill directory. First-time project plumbing: create concepts/, register the skill in agent configs.
• /reframe close → Read and follow commands/close.md in this skill directory. Finalize a concept document with a retrospective and archive it.
• /reframe explain [for <audience>] [about <focus>] → Read and follow commands/explain.md in this skill directory. Translate the concept document into an audience's working vocabulary; output to chat. Doubles as a Phase 7 narration-fidelity comprehension-test artifact.
• No argument → Continue with the methodology below. If concepts/ contains active documents, list them and offer to resume one. If empty, ask the user what they want to reframe and create a new document from references/template.md.

---

Core idea: skeleton vs flesh

Every domain has two layers, and reframing is about which one you rebuild first.

• Skeleton — the load-bearing abstract functions. For a ticket system: information ingestion + state tracking + event notification. Three abstract verbs. Survives every UI redesign for thirty years.
• Flesh — the surface a user sees. Forms, lists, buttons, workflows. Tied to a specific paradigm and era.

Most "AI native" failures rebuild the flesh and call the result innovation. The skeleton stays a CRUD database with stage fields, and the agent becomes a smart input box bolted on top. The reframe never happened.

A real reframe reconstructs the skeleton itself from the new paradigm's primitives. Flesh follows. Often the user-visible result resembles a traditional system — that's fine, that's the projection — but every operation underneath flows through new primitives.

The artifact: concepts/<target>.md

A reframe is a multi-session exploration. The document is the pause-and-resume mechanism: any agent (a future you, a different session, a teammate) can pick up where work left off because the full thinking trail lives in the file, not in any session's memory.

One file per redefinition target. Structure follows references/template.md — the template mandates expressive form (plain-language summary, vignettes, ASCII diagrams, verb-tagged blocks, dependency tables) because abstract methodology applied to abstract content reads like noise without it. Treat the form requirements as defaults; the only documented escape hatch is "v1 skeleton may be prose-only while shape is still forming."

Directory layout

concepts/
  <target>.md          ← active concept document (e.g. ai-native-crm.md)
  archive/
    <target>.md        ← closed concepts moved here by /reframe close

Each file carries YAML frontmatter (target / paradigm / status / opened) so its state is machine-readable. See references/template.md for the exact frontmatter and section structure.

Resume-or-start flow

When the user invokes the skill with no argument:

1. List files in concepts/ (excluding concepts/archive/). For each, read the frontmatter status field; offer files with status: active as resume candidates.
2. If the user picks one, load it and continue from its current phase. If they want a new target, create a new file from the template and start at Phase 1.
3. If no concepts/ directory exists, suggest /reframe init first — but if the user prefers to skip plumbing, just create the directory and document inline.

---

Phase 1: Essence Extraction

Strip the target domain to 3-5 abstract functions that survive any paradigm. Each function is paired with a one-sentence justification and a concrete vignette grounding the abstraction in scene.

Calibration: an essence statement should be specific enough that it could not be implemented the wrong way, and abstract enough that it prescribes no specific implementation.

Write under ## Essence per references/template.md. For failure modes (feature-listing, smuggling paradigm assumptions, over- compression), worked examples, and "when stuck" guidance, see references/phase-guide.md.

Phase 2: Paradigm Primitives

Identify 4-8 operationally concrete building blocks the new paradigm offers. A primitive must be specific enough that a skeleton can be expressed in it; slogans don't count. If a candidate is still a slogan ("AI native"), drill down to operational consequences.

Write under ## Paradigm Primitives. For an example primitive set (AI-native), the drilling-down test, and "when stuck" guidance, see references/phase-guide.md.

Phase 3: Skeleton Reconstruction

Re-express the essence using the primitives — essence expressed in primitives end-to-end, not essence with primitives sprinkled on. Version explicitly: v1 will be wrong somewhere; v2 absorbs what stress tests reveal.

The hardest discipline is refusing the traditional shape when the new shape is unclear. Better to leave a placeholder ("[unresolved in this paradigm]") than to fill it with a familiar mental model.

Output requires a text-diagram of the core flow, a Components → Primitives dependency table, and per-component rationale (v1 may be prose-only). Between consecutive versions, write a delta annotation. See references/template.md for the format and references/phase-guide.md for the "sit with the question" discipline.

Phase 4: Skeleton Stress Test

Probe by removing traditional crutches one at a time. Three possible outcomes per probe: skeleton holds (positive evidence), bends (adjust skeleton, bump version), or collapses (back to Phase 1 or 2 — or document the boundary if the function legitimately stays traditional).

Outcome 3 is not failure. Knowing where the new paradigm doesn't help is part of the design.

Write under ## Stress Tests (append-only) using the verb-tagged block format Date / Removed / Hypothesis / Result / Diagnosis. For probe-question patterns and append-only discipline, see references/phase-guide.md.

Phase 5 (cross-cutting): Transfer Learning

Mature domains have already solved isomorphic problems. Mine for patterns, not shapes. Translate, don't copy. For each candidate source, ask: what abstract problem did it solve, and does the new domain face an isomorphic one? Note deliberate non-transfers too, to prevent re-litigation.

Invoke during Phase 2 (primitive identification), Phase 3 (skeleton drafting), or Phase 4 (rescue when skeleton bends). Write under ## Transfer Log. For example transfers and the two-question test, see references/phase-guide.md.

Phase 6: Flesh

Once the skeleton holds, plan the user-facing surface as a projection of skeleton state. Two governance rules: familiarity is fine, mimicry is suspicious (every flesh element must name the skeleton state it projects); flesh never adds capability the skeleton lacks (if it tries to, revise the skeleton instead).

Write under ## Flesh Plan as a three-column table — element / skeleton state / user need. If you can't fill all three columns for a row, the element isn't flesh — it's old paradigm leaking through. See references/template.md for the format and references/phase-guide.md for the governance rationale.

Phase 7: Comprehension Test

Skeleton soundness ≠ flesh comprehensibility. Phase 7 closes the build-verify loop with the lightest credible evidence that the projection is legible to actual users — symmetric with Phase 4 (where 4 tests bones, 7 tests skin). Without it, the methodology is asymmetric and ships sound skeletons users can't read.

Three diagnoses with three feedback paths when users can't navigate: flesh problem → revise Phase 6; skeleton problem → revise Phase 3 (also the surest signal of "novelty for novelty's sake"); essence problem → revise Phase 1. Misdiagnosis turns the loop into mindless flesh iteration that drifts the skeleton.

Use /reframe explain for <audience> to generate the lowest- fidelity test artifact directly — narration tier. Write under ## Comprehension Tests (append-only) using the verb-tagged format Date / Artifact / Users / Observation / Diagnosis / Feedback. For the four risk tiers, the graduation boundary, and detailed diagnosis rationale, see references/phase-guide.md.

---

Common traps

Five recurring anti-patterns surface across reframes — new shoes on old path, novelty for novelty's sake, premature flesh, shapeless transfer, phantom primitives. Each has a detection method and a cure path back into the relevant phase.

These are diagnostic patterns, not phase-bound. The natural places to check for them are Phase 4 (Stress Test) and Phase 7 (Comprehension Test), where the skeleton and its projection are under test. See references/traps.md for the full catalog.

---

How reframe relates to the other *-driven skills

Reframe is the most upstream of the methodology skills, operating before the system has a settled shape. It pairs with design- driven as predecessor (settled skeleton graduates into design/DESIGN.md via /reframe close), runs parallel to goal-driven on a different axis (shape vs destination), and lightly overlaps evidence-driven at Phases 4 and 7 (claims need falsifiable observations).

For the lifecycle diagram, full pairing details (including when not to substitute design-driven for reframe), and handoff mechanics, see references/cross-skill.md.

When NOT to use this skill

• The domain has clear precedent in the new paradigm (follow the precedent; use design-driven instead).
• The change is incremental within an existing paradigm (design- driven).
• The user wants to describe a paradigm, not design within one (technical-article-writing).
• No specific domain — vague "let's build something AI native" without a target. Get the user to name the domain first; without a target there's no essence to extract.