team-ai-workflows synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A repository-backed knowledge workflow, not an agent runtime
team-ai-workflows is a template for distributing Claude Code skills and team knowledge through one Git repository. This account covers the shipped template at commit 9aebd7a6c2803da450904eb5621ae5439508ff94, with an analysis cutoff of 2026-11-09. SRC-1 supplies inspected shell/Python implementation and model-facing instructions, not an execution trace or experiment. The whole-system boundary includes setup, update-time link maintenance, the Markdown corpus, search and the instructed answer workflow. Claude Code's implementation, actual user installations, external articles and demonstration media do not supply evidence here; see the boundary and Source register.
The division of labor is consequential. Claude Code interprets skills, generates queries, calls tools, reads files and writes answers. Repository shell code exposes skills through filesystem links. Python performs lexical retrieval. There is no locally implemented enclosing agent loop, model provider selection or independent answer evaluator. The host's model version and parameter mutability are unresolved, not evidence of fixed weights or a particular deployed permission boundary (RT-CMP-claude-code).
Installation exposes live content rather than approved snapshots
The prescribed starting path is clone the repository, start Claude Code inside it, invoke /team-setup, then restart the session. Setup asks the host to synchronize user-level skill links, install a Git post-merge hook and persist the repository location in host settings. Confirmation for conflicting personal skills or a different existing hook, and preservation of unrelated settings, are natural-language requirements rather than shell-enforced safeguards (RT-RTE-machine-setup).
The synchronizer creates per-skill references under the user's skill directory. These references expose the current checkout; they do not copy, hash or pin instructions. After a merge, the installed hook invokes the checkout's synchronizer. An edit beneath an existing valid link can therefore change later instructions or executable code without another link-creation decision. Repository customization affords team-authored skills and replacement knowledge, but actual merge review, veto and rollback practice are unspecified (RT-OBJ-skill-links, RT-RTE-repository-admission).
The README's machine-wide distribution claim is conditional, not a demonstrated deployment guarantee. Synchronization keeps any reachable existing entry at a skill name, including a personal directory or another link. Cleanup scans broken entries across the user skill directory rather than only references owned by this repository. Individual creation failures can coexist with the synchronizer's final zero exit, and the hook independently returns zero. A successful status therefore does not establish complete installation. Discovery also depends on the external host and requested restart (RT-RTE-skill-sync, RT-CLM-machine-distribution).
A question drives selective retrieval and a separate file-reading handoff
The representative workflow is /docs-grep, because it is the shipped route from a question to knowledge consumption. It is a source-supported procedure, not an observed successful episode. The user supplies a question; the host is instructed to derive 4–8 keywords, including useful synonyms and language variants, and pass them to Python. This is delegation to a deterministic tool, not to a separate model worker. Direct CLI callers bypass the model query-generation and answer instructions, and Python does not enforce the requested keyword count (RT-RTE-knowledge-search, workflow account).
Python accepts a configured root when its sources child exists; otherwise it falls back to its resolved script location. The configured corpus need not be the repository containing the script, and directory existence does not authenticate its identity or content. The retained location string has routing force, not trust-attestation force (MEM-OBJ-team-workflows-dir).
Within the selected corpus, readable Markdown files are ranked by distinct lowercase keyword substrings matched, then aggregate keyword-line matches, then path order. Duplicate keyword arguments can raise hit counts without supplying independent corroboration. The score measures text overlap, not proposition truth or complete relevance (EPI-RTE-keyword-ranking).
The return contains at most 15 paths, each with scores and at most two 100-character previews from the first matching lines. Those limits constrain emitted search output, not subsequent full-file reads or the host's context window. Truncation and permissive decoding can remove evidence context, while unreadable files are silently skipped. The result is temporary navigation material rather than another durable memory store (MEM-OBJ-search-results, EPI-RTE-preview-output).
The critical next step is separate: the host is instructed to read top-ranked files and answer with their relative paths. The exact number of files and sufficiency of evidence are left to the host. If hits appear weak, it should change keywords and search again. No independent expected answer, fixed threshold, retry budget or final-prose validator is supplied. No-match search returns zero with a retry suggestion; missing corpus configuration returns an error and asks for repair, without performing that repair. Delivery is the subsequent cited host answer, not Python's successful exit (EPI-RTE-weak-retry, EPI-RTE-answer-synthesis, workflow account).
Retained knowledge can guide later work without establishing learning benefit
The strongest supported persistent contribution is the explicit route from editable repository text to later lookup and host guidance. Team members can replace sample documentation or revise instructions; later searches read current corpus files, and valid links expose revised skill content. Corpus prose can retain practical reasons, such as symptom–cause–fix explanations. Those reasons may reach the answering model through requested full-file reads, but previews do not guarantee their inclusion (MEM-OBJ-sources, MEM-BAP-corpus-answer).
These consumers have different authority: retained words affect fixed lexical ranking; retrieved prose supplies reference material; skill text supplies procedural guidance; links and configuration route access. None of these facts turns repository content into verified knowledge or repository instructions into a permission boundary (RT-BAP-user-skills, MEM-BAP-corpus-answer). The README's “repository-shaped RAG” characterization is supported as retrieval code plus requested grounded synthesis, not as a truth guarantee (EPI-CLM-repository-rag). Without an actual source/answer pair, the answer cannot be classified as paraphrase, entailed derivation or ampliative conjecture. Retention and consultation also do not establish evidence-consuming acceptance or integration after acceptance (EPI-RTE-answer-synthesis).
No automatic route from queries, answers, feedback or execution traces to durable behavior-shaping content is shipped in the inspected template. That bounded absence leaves manual experience-derived tips possible and says nothing about external host/provider learning (MEM-ABS-trace-update). Within-question keyword revision is a requested adaptation, not evidence of improved future capacity. Mutable instruction distribution affords later behavioral change, but no retained content-directed criticism episode, subsequent dependent operation or outcome comparison establishes learning benefit, reflective or autonomous theory building, or exercised self-improvement. The runtime does not supply a separate conditions 1–4 assessment from which a theory-builder classification could be inferred; these properties remain independent, not one negative grade (RT-RTE-repository-admission, EPI-RTE-weak-retry).
For the source-native scenario—sharing editable team reference material and consulting it from other projects—the evidence supports the distribution and lexical-selection mechanisms, contingent on setup and host behavior. It cannot establish retrieval completeness, reliable cited answers or improved performance. A deployment trace showing actual discovery, root selection, full-file reads and delivered answers would change the operational assessment. A retained revision episode linked to criticism, later consumption and dependent behavior would address the improvement question; outcome comparisons would still be needed to establish benefit. Independent report verification and profile verification found no blockers or additional declared limits; they do not supply those missing observations.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Source-level mechanisms are not observed deployment | SRC-1; RT-RTE-machine-setup, RT-RTE-skill-sync, RT-CLM-machine-distribution | Shipped setup, synchronizer and hook; no installed state or execution traces | Successful installation, platform compatibility, every-project discovery and safe complete update reconciliation | Installation/update traces with platform, conflicts, errors, resulting links/settings and host discovery |
| Corpus identity, provenance and approval are deployment-dependent | SRC-1; MEM-OBJ-team-workflows-dir, MEM-OBJ-sources, RT-RTE-repository-admission | Configured-root and fallback code plus documented editing; only illustrative corpus supplied | Actual searched corpus/version, trusted real-world API claims, enforced human review and exercised rollback | Deployed root/content identities, provenance, approval records and a documented repair episode |
| Host consumption and context behavior remain external | SRC-1; RT-CMP-claude-code, RT-BAP-user-skills, MEM-BAP-corpus-answer | Skill instructions and tool output, not host internals | Complete model-visible context, instruction precedence, discovery selection, full-file compliance, context eviction or activation guarantees | Host contracts and traces identifying discovered instructions, resolved files, delivered context and consequential use |
| Lexical output and requested citations do not verify an answer | SRC-1; EPI-RTE-keyword-ranking, EPI-RTE-preview-output, EPI-RTE-answer-synthesis | Readable-file ranking/previews and doctrine-only synthesis; no actual answer or acceptance record | Retrieval completeness, source truth, citation correctness, entailed or accepted novel conclusions, and reliable recovery after repeated weak hits | Query/file/answer traces with independently checked evidence coverage, answer criteria and explicit dispositions |
| Persistence and revision affordances do not demonstrate improved future capacity | SRC-1; RT-RTE-repository-admission, EPI-RTE-weak-retry, MEM-ABS-trace-update | Shipped editing, lookup and access-maintenance paths; external host updates excluded | Learning benefit, a retained criticism-driven theory-building loop, or exercised self-improvement | An evidence-responsive revision linked to retained change, later consumer and dependent operation, with outcome comparisons for success |