LLM Project Wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
LLM Project Wiki is a prompt package for creating and maintaining a project-specific Obsidian reference through Claude Code. The reviewed boundary is the complete two-file repository, README.md and SETUP_PROMPT.md, at commit 1fdfba9129a46afd18aa68a83f12d6716ab694c6, with applicability cutoff 2026-10-14. SRC-1 supplies doctrine/design and a README account of example output, not an inspected generated installation, execution trace or experiment. “Whole-system” therefore means the shipped prompt package, not Claude Code or a consuming project. The boundary and Source register fix those distinctions. This synthesis uses the accepted members without additional source reads.
Constructing the reference and its future callers
A user pastes setup into Claude Code at a project root. The foreground model owns the progression: inspect project instructions, memories and structure; answer prerequisite project questions; scaffold a vault; generate and run a source-copying script; build pages domain by domain; wire project rules and commands; log and self-audit. An existing literal Obsidian Vault/ directory instead selects surgical audit and repair. The requested parallel reads do not establish separate agents, isolated contexts or an aggregation protocol. Filesystem and shell effects depend on the excluded host, and no installed permission or isolation envelope is supplied (RT-RTE-vault-bootstrap, RT-RTE-gap-audit, workflow account).
The intended persistent product separates imported Sources snapshots from compiled Wiki pages. Pages contain project descriptions, exact signatures and schemas, non-obvious constraints, cross-references and decision rationales with rejected alternatives. Index navigation, page metadata, templates, schema, a health dashboard and activity/gap logs support later selection and maintenance. These are specified files; Wiki and Sources are also intended to be Git-tracked. Model admission is not per-page human approval: starting setup and later committing its output are distinct from accepting each write (MEM-OBJ-sources, MEM-OBJ-wiki, MEM-OBJ-index, RT-OBJ-schema).
Delivery is the generated installation and support files, not a verified answer set. The final self-audit follows page writing and bootstrap logging, and permits a gap note instead of a completed entity page. It is model-owned checking, not an independent admission barrier or transactional rollback (RT-RTE-vault-bootstrap, workflow account).
Selecting context and repairing knowledge gaps
For a later question or edit, generated project rules request index lookup, relevant Wiki reads, then memory/source fallback when the wiki is insufficient; targeted edits and audits also permit source access. The “Semantic Lookup” table is model-interpreted intent navigation, not a specified embedding search. Requested reads form a pull chain: file content enters the model context by value. No token budget, trimming, isolated worker-return mechanism or compaction/reload protocol is specified. How the host discovers and delivers generated live rules remains uninspected, so these page reads do not establish that every included delivery path is pull-only (RT-RTE-wiki-lookup, MEM-OBJ-index, RT-BAP-project-rules, RT-BAP-maintenance-rules).
The important maintenance distinction is copying versus interpretation. Invoked sync copies project memory, rules and docs into Sources; it does not refresh Wiki pages. Ingest then selects paths using git diff HEAD~1 --name-only, maps references, reads source and page, revises sections, resolves open gaps, creates new pages and logs the operation. Task-time insufficiency can therefore produce a retained one-sentence diagnosis that later guides a durable page revision and subsequent lookup. This is a claimed experience-fed update, not observed learning, raw session extraction or parameter training. The original trace-source category and other project-memory provenance remain unresolved (RT-RTE-source-sync, RT-RTE-diff-ingest, MEM-OBJ-log).
Several conditions limit freshness. Copy errors can be suppressed despite a completion message. A one-comparison diff is not an inventory of all stale state. Dry-run stops before dependent-page grep, so it need not preview the page-level write set. Edit-time stale marking is a documented model instruction, not an inspected hook. The special status consumer falls back when stale, but ordinary lookup has no explicit general stale-rejection gate. Broader audit can compare source inventories and timestamps, inspect actual code, repair pages and fix the script itself; it is a later invoked recovery path, not automatic failure handling. Deletion propagation and obsolete-page withdrawal are unspecified (RT-RTE-source-sync, RT-RTE-diff-ingest, RT-RTE-gap-audit, MEM-CLM-retention-limits).
What checks and repairs establish
Wiki-first rules give pages operational priority, not a truth license or access-control prohibition. The Sources label is a reference-layer policy rather than literal immutability: sync and manual recovery change its files. Warning headers ask the model not to obey three archived instruction copies; they do not establish host instruction isolation and do not cover all Memory/Rules copies (RT-CLM-wiki-first, RT-CLM-sources-immutability).
The design does request source reconciliation, task-sufficiency judgments and structural checks. Their licenses differ. Frontmatter grep checks three key strings, not valid YAML or the required sources field; link/index examples use prefix or substring matches, not complete graph validation. Broader model review could exceed these examples. Neither these diagnostics nor answering without source access establishes semantic correctness. No recorded evidence-consuming epistemic acceptance decision was found in the shipped files; operational admission, retention and cross-linking do not establish accepted new knowledge. Absent source/page pairs also leave copying, entailment and conjecture unresolved for particular page claims (EPI-RTE-schema-check, EPI-RTE-link-check, EPI-OBJ-wiki-pages, EPI-CLM-answer-coverage, epistemic bounded conclusion).
The strongest improvement-related contribution is a specified repair pathway for the generated installation's own behavior-guiding files. Audit diagnoses incomplete copying or maintenance wiring and can revise scripts, commands, schema and project rules for future sessions. Its independent objective is complete, fresh coverage and closed maintenance wiring across sessions. Reflection and improvement-directed self-change are claimed, not exercised: no generated revision, live loading or subsequent change-dependent behavior closes the causal chain. Successful improvement is unestablished, and this pathway does not revise the shipped setup prompt (RT-RTE-gap-audit, RT-BAP-project-rules, RT-BAP-maintenance-rules).
The theory-builder conditions remain independent. Condition 1, localized maintenance-rule content, and condition 2, decisions depending on that content, have claimed status. Condition 3, content-directed criticism naming claim, result and blame, and condition 4, criticism-dependent iteration, remain uninspected. Structural repair does not itself establish criticism of the wiki-first premise. Whether criticism of a consumed theory improved capacity for future action is likewise uninspected. Reflective and autonomous theory-builder status are uninspected; computational patching alone does not establish every internal theory-building operation. These uncertainties do not negate the specified experience-fed page repair or method-file self-change (RT-RTE-gap-audit).
For the scenario of a project session answering from maintained pages, the design supplies a coherent selection and recovery procedure, conditional on accurate compilation, metadata, maintenance invocation and host compliance. For a scenario requiring reliably fresh or warranted answers, those conditions are not demonstrated. An installed vault with source/page comparisons would resolve compilation and coverage questions; traces linking gaps to repairs and later consumers would establish operation; an independently assessed before/after comparison would be needed to attribute better future capacity. Inspectable edit hooks, complete dependency reconciliation and explicit withdrawal rules would materially change freshness guarantees. Current evidence supports intended organization and use, not a product ranking or demonstrated benefit (EPI-CLM-compiled-knowledge, RT-CLM-audit-completeness).
Limitations
The accepted report verification and profile verification declare no residual verification limits. No unresolved cross-report conflict or additional record fault was found. The following evidence and specification limits still bound the account.
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Design without generated installation or run evidence | SRC-1; RT-RTE-vault-bootstrap, RT-RTE-wiki-lookup | Complete two-file prompt package; README example is reported operation | Successful generation, persistence/reload, actual source-read reduction or answer benefit | Generated artifacts and linked bootstrap/later-session traces; outcome comparisons for benefit |
| Host, provider and plugin behavior excluded | SRC-1; RT-CMP-claude-code, RT-BAP-project-rules, MEM-OBJ-health | Specified host channels and plugin uses, not their implementations | Model-specific reproducibility, permissions/isolation, guaranteed instruction loading, complete delivery-direction classification or query execution | Pinned host/model configuration, loading traces and concrete plugin/query implementation |
| Initial provenance and concrete source coverage unknown | SRC-1; MEM-OBJ-sources, RT-RTE-source-sync | Standard memory/rules/docs copy pattern and allowed broader project inspection | Exhaustive entity acquisition, complete lineage or initial admission-control classification | Installation-specific input provenance, adapted script and project/source inventory comparison |
| Page semantics and warrant unestablished | SRC-1; EPI-OBJ-wiki-pages, EPI-CLM-answer-coverage | Required page content, quality criteria and model-directed admission | Definite compilation semantics, checked derivations, accepted ampliative claims or correct answer coverage | Source/page pairs, semantic checks and retained acceptance decisions with criterion, use and scope |
| Freshness and withdrawal incomplete | SRC-1; MEM-OBJ-page-metadata, MEM-OBJ-log, RT-RTE-diff-ingest, MEM-CLM-retention-limits | Documented grep/diff/gap/stale chain; no deletion protocol | All stale pages refreshed or withdrawn; deleted entities excluded; complete dry-run write preview | Consistent dependency paths, status-reset/closed-gap semantics, deletion and retirement rules, broader change reconciliation and exercised traces |
| Diagnostics narrower than completeness claims | SRC-1; EPI-RTE-schema-check, EPI-RTE-link-check | Embedded key-presence and navigation examples | Full schema validity, exact graph connectivity or semantic correctness | Concrete parsers/checks and results covering required fields, exact links and content claims |
| Experience-fed update origin and effect unresolved | SRC-1; MEM-OBJ-log, RT-RTE-diff-ingest | One-sentence task-gap reports and prescribed later page repair | Controlled original trace-source classification, observed trace-fed updating or improved capacity | Original activity records linked to gap diagnoses, page changes, later consumption and evaluated outcomes |
| Method repair does not close criticism or self-improvement evidence | SRC-1; RT-RTE-gap-audit, RT-BAP-maintenance-rules | Specified diagnosis, method-file patching and future instruction channels | Theory-builder conditions 3–4, autonomous theory building, exercised self-improvement or successful learning | Retained claim/result/blame criticism shaping another round; update, live authority and subsequent behavior traces; improvement comparison against the independent objective |