ISA

• Every ISA the skill writes or touches opens with a ladder that passes the shape check, is written from the summit, and agrees with its criteria; anything in the ISA that disagrees with the ladder is resolved before another criterion is climbed.
• L3 and L2 reach the principal through Restate before a new project or a changed direction is built (an extension whose re-derived L3 or L2 changed counts as one), and not otherwise.
• An existing ISA is extended in place, never replaced by a second ISA: the new work takes the next F<n> block (or new flat criteria when the ISA is flat) and the next global ISC IDs, no existing ID renumbers, the ladder is re-derived so it covers the new work, and when the re-derived L3 or L2 changed, they reach the principal through Restate before the new work is built.
• Every criterion is atomic, names its falsifier, and carries at least one ## Test Strategy row in the canonical six-column shape, so Assay resolves the whole table with no dropped row. Each row runs on a lane, liveness for asserts on every monitor tick or integrity for checks that may judge (or build, full runs only), and at a stage, so a criterion's asserts run before its judged rows (ISAFormat § Verifier classes, lanes, and stages; the one progressive contract is skills/Assay/References/Progressive.md).
• Discuss is the standing alignment loop: whenever the ISA's state says direction is uncertain (bun LIFEOS/TOOLS/IsaAlign.ts status <isa> returns DISCUSS) or the conversation shows doubt about what done means, a Discuss round runs before the uncertain part is built, sized quick, full or deep to the doubt, and Restate, Refine, Rate and the ambiguity check hand off to it when their single-question shapes stop being enough.
• Every interviewing surface (Discuss, Interview, Questionnaire, Refine, Scaffold's ambiguity check) leaves its raw record in the ISA's append-only ## Articulation Log, and the settled answers reach the ladder and the sections in the same turn.
• The ISA's history stays auditable: criterion IDs never renumber, and Decisions, Learning and Verification entries keep their canonical shapes.

Before executing an ISA operation, load its operation bundle:

bun "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/LIFEOS/TOOLS/ContextLoad.ts" ISA load <bundle>

The bundle returns required ISA constraints, operation instructions, and internal references with source hashes. Follow every returned requirement, including conditional reads. Skill-local paths resolve from the skill root; LIFEOS/ paths resolve from the install root. Do not reread unchanged context still present; after compaction or a source change, reload the relevant bundle.

Set the shell and any code-mode wrapper output budget to 30000 tokens. Treat any truncation warning as a failed load, even when END CONTEXTLOAD appears. Use list to discover bundles and search "<query>" to search the complete guide. A failed search does not prove absence: inspect the index, broaden the query, or load full. If routing is uncertain, a bundle is absent, or the loader fails, load or read References/CompleteGuide.md in full before acting.

When the verb is ambiguous, default to Scaffold for a new ISA and CheckCompleteness for an audit.

The format authority is ISAFormat and References/CompleteGuide.md; this skill takes no install config. <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ISA/History.md, when present, holds the install-specific incident detail behind the dated gotchas here and in the intents; it is history, never read as config.

These three bite every intent, so they live here; the rest are in References/CompleteGuide.md § Gotchas.

• Test Strategy tables must carry the canonical six columns, isc | type | check | threshold | tool | anchors_to. The shipped parser (LIFEOS/TOOLS/lib/IsaTestStrategy.ts, which ISAGate.ts runs) maps cells by the header row and falls back to these canonical positions (anchors_to at cell index 5) for any column the header does not name, so a row whose cells do not line up with its header shifts the anchor into the wrong slot; ISAGate.ts then hard-fails anchors-missing on rows that visibly show an anchor, and the message points at the anchors rather than at the column count. Drop threshold only by writing an empty cell, never by removing the column. (Hit live 2026-07-27 on a large ISA where every flagged row was anchored; incident detail in the USER extension History.md.)
• The Articulation Log tolerates no embedded headings. Inside a log block, no line may start with #: demote pasted or transcribed headings to bold or blockquote them (> ## Original). CRITERIA_HEADING_RE accepts ### Criteria and parseCapabilities opens on H3 Capabilities, so on a criteria-less ISA (exactly what Discuss creates before Scaffold runs) an unquoted transcript heading becomes the first parser match and everything after it parses as criteria. Only the block's own ### YYYY-MM-DD HH:MM · <source> headers are legal.
• Browser/HTML view of an ISA goes through bun LIFEOS/TOOLS/ISARender.ts <slug-or-path> [--output <path>], never a hand-rolled renderer or the generic HTML skill. It emits the canonical technical ISA template, deterministic and offline, with per-ISC status pills, the ascent strip, and the Test Strategy table. (2026-08-31: an ISA was hand-rendered twice, first as a distilled report, then as generic markdown-to-HTML, because this tool was referenced nowhere in the skill; the principal's correction encoded here.)

User: "Scaffold an ISA for the new billing dashboard"
→ Scaffold intent (bundle `scaffold`)
→ A populated ISA at its canonical path, ladder first, every criterion with a Test Strategy row, and L3 and L2 shown through Restate before the build starts

User: "Extend the Lighthouse ISA with a new Feature F7, then restate L3/L2"
→ Scaffold intent (bundle `scaffold`), extending the existing file in place
→ The same ISA with F7 added under the next global ISC IDs, nothing renumbered, the ladder re-derived, and the amended L3 and L2 restated

Restate answer: "sort of, but it is really for teams, not solo users"
→ Discuss intent (bundle `discuss`), align mode, quick depth
→ The ladder shown as it stands, one round of up to four recommended-answer questions including a challenge, the round checkpointed, L1/L2/L3 and the criteria folded, then the corrected L3 and L2 restated

User: "Is this ISA ready to close?"
→ CheckCompleteness intent (bundle `check`)
→ A structured gap report: each gap names its section, severity and reason, and `status: pass` only with zero hard gaps