scitexification

name: scitexification description: | [WHAT] Scitexification — the verb: translate existing code (a script, a notebook, a small repo, a published-paper supplement) into SciTeX form so it gains the ecosystem's session-managed I/O, scitex-clew evidence binding, publication-quality figures, and project structure. Single source of truth for the migration act itself, package-agnostic; per-package patterns delegate to the per-pkg SKILL.md, and Clew-specific specialization layers on top of this skill via scientific/04_clew_*. [WHEN] You have existing code that works and want to bring it into the SciTeX ecosystem with minimal rewrite, OR an agent is asked to translate a research capsule, paper supplement, or one-off notebook into a SciTeX project. Load this skill BEFORE picking the per-chapter topic. [HOW] Read this SKILL.md first to orient on the 5-stage translation arc; then drill into the topic chapter you need next (01_io-patterns, 02_session-config, 03_plt-patterns, 04_repro-clew, 05_naming-and-numbering). Per-package details (full stx.io API surface, figrecipe figure types, etc.) live in the corresponding per-pkg SKILL.md — this skill references them rather than restating. tags: [scitexification] requires:

Scitexification is the translation act; the four package-level

companions supply the API knowledge an agent needs to actually

translate-and-resolve. Loading `scitexification` should also

surface these. See 00_playbook.md §"Required companion skills".

scitex-session
scitex-io
figrecipe
scitex-clew user-invocable: true primary_interface: skills interfaces: python: 0-3 cli: 0-3 mcp: 0-3 skills: 3 http: 0

Scitexification — the translation act

pip install scitex-dev (no extras needed for the skill itself).

scitexify (verb, derived) — to convert existing research code into the SciTeX idiom: session-managed I/O, project-structured outputs, evidence- bound claims, publication-quality figures, mirrored test layout.

This skill is the single source of truth for the translation act. It is deliberately package-agnostic at the top level; each chapter delegates package-specific patterns to the corresponding per-pkg SKILL.md so the per- package teams own the surface area they ship.

Canonical universal playbook — 00_playbook.md. Read this first when scitexifying any artefact. It coins the vocabulary (scitexify / scitexification / scitexified), fixes the universal contract (universal inputs, pre-flight, phase dispatch, done condition, forbidden), and frames the honest source-grounding principle — "attempt every claim, ground where possible, and when ungroundable include the claim with null + a reason, NEVER silently omit" — as a general scientific-integrity norm independent of any specific evaluator. The per-chapter files (01–05) drill into stage-specific patterns; the 04_clew_* siblings are clew-tracked specialisations that compose on top of the universal playbook.

When to load this skill

Load when any of the following is true:

You inherited working code (script, notebook, small repo) and want it inside the ecosystem with minimal rewrite.
An agent is asked to translate a research capsule, paper supplement, or one-off notebook into a SciTeX project.
You are about to hand-write data/results/claims.json (or submission.json, etc.) — stop, read chapter 04_repro-clew, use the API.
You are about to copy-paste a matplotlib figure call from a paper template — stop, read chapter 03_plt-patterns, use the figrecipe pattern.

Do not load this skill when:

You are starting a brand-new SciTeX project from scratch — go to ../02_research-project_* (project structure rules).
You are auditing or building a SciTeX package (i.e. publishing scitex-xxx) — go to ../../general/ (engineering rules for package authors).

The 5-stage translation arc

Scitexification is not a search-and-replace. It is five staged transforms each of which holds independently:

Stage	Chapter	What changes	What stays the same
1	01_io-patterns	Every `open()` / `np.load` / `pd.read_csv` / `pickle.load` becomes `stx.io.load(...)`; every `np.save` / `pickle.dump` / `df.to_csv` becomes `stx.io.save(..., symlink_to=...)`. DAG composition (output of step N is input of step N+1) becomes visible at the filesystem level.	Your algorithm. Your data shapes. Your business logic.
2	02_session-config	The script entry-point becomes `@stx.session.start(...)`; magic numbers and paths become `CONFIG.<KEY>` lookups against `config/*.yaml`; logging becomes the session logger.	Function call structure. Module organization. Test cases.
3	03_plt-patterns	Every `plt.savefig(...)` becomes a `stx.io.save(fig, ...)` (so the figure is bound to a session output), and every visual style choice ladders up to figrecipe's publication-quality primitives.	Figure intent (what comparison, what axis labels). What information the figure carries.
4	04_repro-clew	Final-mile assertions (`accuracy was X%`, `effect size was Y`) become registered Clew claims, each evidence-bound to the file that produced it; the submission JSON is composed by iterating registered claims through `scitex_clew.list_claims()` + `scitex_clew.verify_claim()` and filtering to `source_verified=True`, not hand-written.	What you are claiming. Your numbers.
5	05_naming-and-numbering	`cnn_v3_final_FIXED2.py` becomes `scripts/03_cnn.py` (zero-filled, sortable, mirrored under `tests/`); IDs and ordinals become readable symlinks per `02_research-project_09`.	Your filenames as a concept. The numbers themselves (after zero-fill).

Doing stages 1+2 alone gets you a runnable SciTeX project — stage 3+ are strictly additive. If a capsule's deadline is tight, stages 1+2 are the minimum viable scitexification; stages 3, 4, 5 land in subsequent PRs.

Relationship to other skills

This skill does not duplicate content elsewhere; it composes them.

For project structure (where files go, what ./config/ looks like, what ./data/ allows): see ../02_research-project_*. This skill assumes a working knowledge of that structure as the target of the translation.
For per-package API surface (the full stx.io save/load type matrix, figrecipe's figure types, scitex-clew's primitive operations): see the per-pkg SKILL.md (~/.claude/skills/scitex/scitex-io/, .../figrecipe/, .../scitex-clew/). This skill teaches which primitive to reach for during translation, not what the primitive does internally.
For Clew-specific translation (capsule-aware DAG, oracle-blind scoring, DONE signalling): see 04_clew_* skills. Those are specializations of scitexification stages 1+2+4 for the Clew-tracked cohort flow. If you only need to scitexify and don't need Clew verifiability, ignore the 04_clew_* skills.
For PDF reporting (recurring scientific PDF deliverables): see 03_reporting_*. Reporting is a downstream concern; scitexify first, then report.

Tags & discovery

This skill uses a parent + sub-tag scheme:

tags: [scitexification]              # the umbrella tag — include all 5 chapters
tags: [scitexification, scitexification.io]      # chapter 01 only
tags: [scitexification, scitexification.session] # chapter 02 only
tags: [scitexification, scitexification.plt]     # chapter 03 only
tags: [scitexification, scitexification.clew]    # chapter 04 only
tags: [scitexification, scitexification.naming]  # chapter 05 only

Capsule-side declarative interface (target, gated on the scitex-dev / SAC discovery contract — see A2A thread 48d2324b):

# <project-root>/.scitex/dev/config.yaml
skills:
  tags: [scitexification]            # full series
  # or, for a stripped-down subset:
  tags: [scitexification.io, scitexification.session]

Until the declarative-discovery contract is finalized, an agent can manually expand the tag:

scitex-dev skills tags-expand scitexification

and read the printed paths. The cohort-A reference consumer (proj-paper-scitex-clew) drives the prompt-side migration plan: their bespoke PROMPT_SCITEX_TRANSLATION_FOR_CLEW.md will be retired once this skill ships and is wired through the discovery contract.

Pitfalls (the migration traps)

Patterns that look "almost SciTeX" but aren't, and that this skill exists to keep you out of:

Hand-writing claims.json / submission.json even though the project has Clew registered. The hand-written JSON drifts from the evidence-binding the registered claims actually have. Always compose from scitex_clew.list_claims() + filter on scitex_clew.verify_claim(c).source_verified — see chapter 04.
Calling plt.savefig from a SciTeX session script because "matplotlib already wrote the file." The file lands outside the session's output dir, gets timestamped against the wrong run, and silently breaks make repro. Always wrap in stx.io.save(fig, ...) — see chapter 03.
Mixing os.path.join(...) and CONFIG.PATH.<KEY> inside the same script. Either fully translate to the CONFIG lookup or stay fully on os.path.join for the call site; mixing is the loudest tell that scitexification stage 2 was rushed. See chapter 02.
Renaming final_v3_FIXED.py to script_final_v3_FIXED.py under the impression that putting it in scripts/ is enough. Stage 5 exists because filename hygiene is load-bearing for downstream reproducibility — see chapter 05.
Skipping tests/ mirroring because "the script already works." See ../02_research-project_06_project-structure-tests.md; the mirroring is what lets audit-project catch drift.

Status (2026-06-04 draft)

✅ This overview SKILL.md drafted.
⏳ Chapters 01–05 queued for drafting, pending sign-off from operator + proj-paper-scitex-clew on this overview.
⏳ Parent + sub-tag discovery syntax (scitexification vs scitexification.io) requires a small extension in scitex-dev/_cli/skills/_tags.py — separate issue to be filed.
✅ Chapter 04 uses scitex-clew primitives (list_claims, verify_claim, render_dag) rather than proposing a new emit_submission API. Rationale: scitex-clew stays general-purpose; the submission shape (which keys, which schema) is cohort/paper- specific concern, owned by the consumer (paper-scitex-clew, MNIST template, etc.) as a 10-line iterate+filter+emit helper. Per operator decision (Telegram msg 125).
⏳ Capsule-side discovery contract (spec.claude.skills: [scitexification] auto-loading) tracked in A2A thread 48d2324b between proj-paper-scitex-clew ↔ proj-scitex-dev.

References

~/.claude/skills/scitex/scitex-io/SKILL.md — full stx.io API surface.
~/.claude/skills/scitex/scitex-session/SKILL.md — @stx.session.start contract + CONFIG mechanics.
~/.claude/skills/scitex/scitex-clew/SKILL.md — clew primitives, registration, verification.
~/.claude/skills/scitex/figrecipe/SKILL.md — figrecipe figure types, publication-quality defaults.
~/.claude/skills/scitex/scitex-template/SKILL.md — the cookiecutter research-project scaffold (the target shape of scitexification).
../02_research-project_* — what the scitexified result should look like structurally.
../04_clew_* — Clew-tracked specialization of stages 1+2+4.
../05_private-skills_01_consumer-project.md — where consumer-project-private notes live (post-scitexification).