Interview Doc Agent synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Interview Doc Agent is a prompt-operated experience library and document workflow, not an enclosing agent runtime. It asks an external host AI to organize personal project records into reusable dossiers, maintain lookup indexes, and later generate resumes or interview scripts from selected material. This account covers the complete shipped prompt-and-document artifact at commit 6ebfe9aba989d7949dadea0fc4af6fb64892c6aa, with an analysis cutoff of 2026-10-14. Its evidence is exclusively doctrine/design: inspected instructions, tutorials, templates and fictional examples. It contains neither an inspected execution engine nor a supplied real input/output trace or comparative experiment. External hosts, model providers and linked sources are excluded. See the boundary and evidence and Source register.
From raw experience to retained dossiers
The user supplies both the materials and the requests. A host such as Claude Code, Claudian or a Feishu assistant supplies model computation, conversation, permissions and file/page tools. The model interprets natural-language guidance and owns the next step; the artifact does not supply its own scheduler, dispatcher or independently enforced tool policy. Initialization requests a directory layout, blank profile and lookup index. It does not create populated autobiographical memory or verify that later reads resolve the same workspace. These roles are documented in SRC-1, SKILL.md: RT-CMP-host-ai, RT-OBJ-skill, RT-RTE-initialize-workspace and RT-BAP-skill-instructions.
The central handoff is raw records → project dossier → updated indexes → a later document request. On a user's organization request, the model is directed to extract background, problems, actions, results, why/how explanations, quick-access figures and persona strengths. It writes the corresponding dossier and updates both the personal profile's project index and wiki/index.md. Raw originals remain reachable. In the local procedure, the model admits the transformed content without mandatory per-content human approval; the user's initiation is not such approval. This is a declared persistent content-update route, not demonstrated successful extraction or improved capability (SRC-1, SKILL.md, materials/README.md; RT-RTE-consolidate-experience, MEM-OBJ-materials, MEM-OBJ-experience-dossiers).
The two access surfaces have different content. The wiki table provides paths, keywords and quick figures. The personal profile combines biography and persona content with project locators. Both are text interpreted by the model, not a shipped fixed-rule retrieval schema. Dossiers preserve reasons for the user's project decisions; that does not establish retained reasons for the model's extraction or claim-level source provenance (SRC-1, wiki/index.md, 经历库/00-个人档案.模板.md, 经历库/01-示例经历.md; MEM-OBJ-wiki-index, MEM-OBJ-personal-profile, MEM-OBJ-experience-dossiers).
From retained context to a delivered document
The representative workflow is user-requested consolidation followed by a separate request for a JD-tailored resume. The model reads the lookup table, profile, relevant whole dossiers and a writing template. A supplied job description guides relevance, reordering and emphasis, not factual authentication. Resume guidance compresses conclusions and quantities; interview scripts expose project reasoning. If a dossier lacks detail, the model may return to raw material. Thus the library is the primary factual source, not an exclusive source (SRC-1, SKILL.md, jd/README.md; RT-RTE-generate-document, MEM-OBJ-jd, MEM-BAP-template-guidance, workflow account).
Selection serves the foreground document request. It can reduce requested reads but does not isolate a context, evict prior turns, restrict access to unread files or impose a fixed token budget. The rationale also allows reading the entire small library. Neither whole-file access nor possible window fit demonstrates complete delivery, accurate selection, token savings or the advertised 100% retrieval reliability (SRC-1, docs/01-理念-karpathy-wiki.md; RT-RTE-generate-document, MEM-CLM-retrieval-reliability).
The missing-number policy is a substantive declared constraint: preserve quantities, do not invent data, and use ___ with a reminder when genuine measurements are missing. It withholds a number rather than necessarily stopping delivery. Its force is model-interpreted policy, not an independent numerical checker. The included dossier explicitly contains fictional demonstration data, so storage in the experience library cannot authenticate a fact. Common-source selection likewise cannot guarantee agreement across rewritten documents, raw-detail additions, source updates and human edits (SRC-1, SKILL.md, 经历库/01-示例经历.md, docs/04-使用指南.md; RT-CLM-numeric-fidelity, EPI-CLM-source-consistency).
Within this shipped artifact, no separate evidence-producing truth check, evidence-consuming acceptance decision or subsequent acceptance-to-integration route is established. Retention and reuse give candidate assertions operational influence, not independently established factual warrant. Without real input/output pairs, the account also cannot decide whether rewritten explanations merely preserve meaning, entail new propositions or introduce non-entailed conjectures. This does not deny possible private model comparisons or later human review (SRC-1, SKILL.md; EPI-ABS-content-check, epistemic ledger).
Local delivery means saving a requested document under output/ for the user to take and adjust. It does not mean factual certification, PDF compilation or employer submission. In the Feishu adaptation, automatic writeback depends on permissions; denied writes lead to conversational text that the user can manually paste. That is a distinct human admission route, not a factual acceptance test or general partial-write recovery protocol. The cloud host's later selection and transformations are uninspected, so local whole-file delivery cannot be attributed to it (SRC-1, docs/04-使用指南.md, docs/03-接入-飞书.md; MEM-OBJ-output, RT-RTE-feishu-writeback).
What the progression establishes
For repeated job-document requests over a personal library, the strongest supported contribution is a declared route that retains organized experience and reuses it through distinct factual, style and relevance channels. Request-triggered revision of dossiers and indexes is also specified. Neither contribution establishes improved capacity for future action: no performance comparison or later activation evidence is supplied. Automatic derivation and later consumption are described, but the chats and retrospectives accepted as inputs do not establish a qualifying trace origin. Trace learning is therefore unresolved beyond bounded local negatives, not absent system-wide (MEM-BAP-dossier-context, RT-RTE-consolidate-experience, MEM-ABS-memory-maintenance, profile verification).
Reusable why/how accounts do not establish theory building. The accepted runtime does not supply separate supported theory-builder condition 1–4 assessments; this synthesis therefore makes no positive or blanket negative finding on those conditions. Criticism of a consumed theory improving future capacity is unestablished. Reflection, autonomous theory building and self-improvement are independently unestablished, not a collective failure grade. The shipped workflow is user-triggered rather than autonomous end-to-end scheduling; no performance-fed revision of the skill or templates is prescribed. Generated products have a human consumer, not a prescribed route back into method guidance (runtime account, MEM-ABS-memory-maintenance, MEM-OBJ-output, epistemic conclusion).
A real consolidation-to-generation trace with source comparisons would change the assessment of execution and semantic fidelity. Retrieval reference sets and token measurements would address recall and efficiency claims. An evidence-consuming factual gate would change the declared acceptance architecture. A retained, evaluated method update followed by dependent later operation would be needed to assess improvement through method criticism. None is supplied here. Local file placement also leaves provider transmission unresolved; the cloud variant is expressly cloud-based (RT-CLM-local-privacy).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Workspace root and starter-file names do not resolve uniformly in the instructions; default-template fallback remains unresolved. | SRC-1; RT-RTE-initialize-workspace, RT-RTE-generate-document; report-verification limit | SKILL.md, local installation tutorial, profile and template naming at the frozen commit |
Guaranteed out-of-the-box discovery, resolved fallback and uniform cross-host setup | An explicit root/name-resolution contract and inspected deployment traces covering these branches |
| All behavior evidence is doctrine/design; examples are fictional and outputs/traces/experiments are absent. | SRC-1; RT-RTE-consolidate-experience, RT-RTE-generate-document, MEM-OBJ-output | Complete shipped artifact, excluding external host implementations | Successful writes, cross-session activation, reliable execution, output quality or causal benefit | Real input/output traces with write/read-back evidence and appropriately scoped comparisons |
| Source authenticity, claim-level lineage and semantic preservation are unresolved; policy is not independent checking. | SRC-1; MEM-OBJ-experience-dossiers, RT-CLM-numeric-fidelity, EPI-ABS-content-check | Raw-to-dossier-to-output instructions and example artifacts | Accepted factual knowledge, checked causal explanation, guaranteed fidelity or across-output agreement | Authenticated source material, aligned transformations and evidence-consuming checks with consequential acceptance decisions |
| Selective reads have no enforced context budget or delivery/recall measurement. | SRC-1; RT-RTE-generate-document, MEM-CLM-retrieval-reliability | Local full-file/index-guided instructions, including raw fallback and whole-library alternative | Universal recall, context isolation, measured savings or superior inference | Host-visible selected payloads, relevance references and controlled token/quality comparisons |
| Cloud intermediaries, actual permissions, selector and consumed representation are uninspected. | SRC-1; RT-RTE-feishu-writeback | Feishu tutorial and manual-copy alternative; host internals excluded | Uniform local/cloud memory classification, full-page delivery, pull-only system behavior or synchronization guarantees | Host integration and retrieval evidence tracing persisted content through selection to later consumption |
| Initial populated profile control is unspecified; maintenance lacks a shipped withdrawal, transaction or rollback protocol. | SRC-1; MEM-OBJ-personal-profile, MEM-ABS-memory-maintenance | Profile initialization and user-requested dossier/index refresh | Complete initialization lineage/agency, reliable multi-file consistency or withdrawal of obsolete output claims | Explicit population/admission roles and implemented, traced consistency/withdrawal/recovery procedures |
Trace-learning aggregate coverage is overcautiously labeled not-determinable despite a classified local negative; dossier trace origin and cloud alternatives remain unresolved. |
SRC-1; MEM-ABS-memory-maintenance, RT-RTE-initialize-workspace, MEM-OBJ-output, MEM-OBJ-materials, RT-RTE-consolidate-experience, RT-RTE-feishu-writeback; profile-verification limit | Classified manual/blank/human-product routes alongside heterogeneous dossier inputs and opaque cloud processing | A system-wide trace-learning negative, or interpreting the aggregate as establishing no local controlled value | Preserve unit-level negatives; partial is the more precise coverage label. Original-input provenance and cloud update/consumer evidence would resolve remaining units |
| Local storage does not determine provider transmission; personal-data ignore patterns are commented out. | SRC-1; RT-CLM-local-privacy | Local-installation documentation and ignore rules; provider transport excluded | No-upload privacy guarantee or automatic exclusion of personal data from Git | Actual host/provider data-handling contracts, transport evidence and active repository exclusions |