Skip to content

Knowledge: Process

Topic file for process facts. Updated by /tfw-knowledge. See KNOWLEDGE.md §4 for the index.

# Fact Verified Source(s) Added
F1 TFW = teamwork. The framework's purpose is to enable collaboration between multiple agents and humans ⚠️ 1 source Chat TFW-18 (user) 2026-04-03
F2 The best fact candidates come from user emotions and frustration in chat, not from structured artifact sections. When the user is angry, excited, or sharing insights during research — that's the primary signal for strategic knowledge ⚠️ 1 source REVIEW-18B FC#2, chat TFW-18B 2026-04-03
F3 Naming creates behavior in AI agents more effectively than explanation. Right terminology triggers right associations. «Маленький промпт + точные термины > длинный промпт с объяснениями». Claude «dreaming» = 1 word replacing paragraphs. TFW adopted: OODA, Sufficiency Verdict, Trust Protocol, Progressive Disclosure ✅ verified RES TFW-22 FC#2, REVIEW TFW-22 FC#2 2026-04-04
F4 AI agents follow numbered steps + gates perfectly but lose focus in prose workflows. Workflow = algorithm (steps + gates + refs), not info dump. Observed: executor doesn't deviate from TS steps. Confirmed across TFW-22 and prior tasks ✅ verified RES TFW-22 FC#3, REVIEW TFW-22 FC#3 2026-04-04
F5 AI should write findings to file FIRST, then summary to chat. Protects against session interruption/death. Adds ~10-15% tokens, not 100% ✅ verified RES TFW-22 FC#5, REVIEW TFW-22 FC#6 2026-04-04
F6 Coordinator without visionary oversight drifts into scope explosion. Phase A coordinator added tasks/, reference resolver, workflows/templates — all beyond master HL. Need lead coordinator with veto authority across sessions, or at minimum strong HL constraints ⚠️ 1 source TFW-26 coordinator review session 2026-04-05
F7 Agents in different chat sessions lose strategic decisions context. User had to re-explain coordinator decisions from one chat to another. §11 Strategic Insights helps, but information loss between sessions is a systemic TFW problem ⚠️ 1 source User, TFW-26 session: "пришлось заново объяснить координатору" 2026-04-05
F8 Executor role lacks explicit session fact capture step. Coordinator has §11 in HL template + plan.md Step 4b trigger. Executor sees user reactions live during testing but has no mechanism to record them. Insights are lost without explicit capture ⚠️ 1 source RF TFW-26/B obs. #9, executor self-report 2026-04-05
F9 CTA (star, share, recommend) goes after value delivery, not during onboarding. User gets value first → then call-to-action. Standard marketing pattern applied to TFW init: star recommendation at Phase 5 Verify, after project is set up and first task closed ⚠️ 1 source User insight, TFW-31 (marketing timing) 2026-04-09
F10 User's actual TFW init pattern: point agent at .tfw/ folder + describe business context in natural language. «Мы начинаем новый проект. [контекст]. Сделай инициализацию.» Agent gets file access + business context + instruction to initialize ⚠️ 1 source RES TFW-31 FC3 (user interview) 2026-04-09
F11 Organic emergence → formalization = reliable pattern for discovering workflow improvements. If agents invent a pattern independently across iterations (e.g., Iteration Status + Open Threads in TFW-32 RES3-4), it should be codified ⚠️ 1 source REVIEW TFW-32/C FC#1 2026-04-10
F12 Analytical phases (positioning specs, improvement direction) have structurally different deliverables — no code, no tests, no modified files. The 9-point review checklist's code-centric items all evaluate to N/A. A lighter review template for analytical phases could reduce friction ⚠️ 1 source REVIEW TFW-32/D FC#2 2026-04-10
F13 Every research iteration needs a SPECIFIC trigger — «just run another iteration» without trigger = waste. Trigger taxonomy: error correction, gap filling, depth on specific finding, user-injected new direction. Observed across 6 iterations (VLM-3: 4, TFW-32: 2) ⚠️ 1 source RES3 FC15 (cross-project pattern) 2026-04-10
F14 Without YAML control files or explicit statuses, agents fast-run every time. «Agents always want to finish faster.» Structural enforcement (YAML + gate + min_iterations) is required for any iterative AI workflow ⚠️ 1 source RES2 SS2 (user principle) 2026-04-10
F15 Section names in templates = prompts for writers, not tags for collectors. Collection algorithm reads full artifacts — doesn't need standardized names to find content. «не проблема заставить агента заглянуть в каждый раздел» ⚠️ 1 source RES4 SS10 (user insight) 2026-04-10
F16 AI-generated citations propagate through TFW role transitions without verification — "fabrication chain." In TFW-36 Phase A: (1) "55 verified facts" (actual: 18 verified of 55 total), (2) "0.38% context quality" attributed to Anthropic (fabricated — real 0.38% is about jailbreak refusal rates), (3) "85% AI projects fail" attributed to McKinsey (actual source: Gartner 2019). All three traversed Researcher → Coordinator → Executor → Reviewer (4 roles, 8+ documents). Code tasks have compilers to catch this; content tasks need a Source Audit gate — mandatory citation register with verified URLs before RF status ⚠️ 1 source TFW-36 Phase A incidents, REVIEW §2 2026-04-13
F17 Template-based state file creation came from user insight, not framework analysis. User suggested "template for init" as the solution to knowledge_state.yaml contamination — before any structured investigation. Validates P2 (process.md): best insights come from user's direct experience with the problem ⚠️ 1 source HL TFW-40 §11 S3 (user, initial report) 2026-04-15
F18 When TS ACs target an existing template for modification, section numbering drifts if ACs reference sections by number ("§4 should be AC") and a later AC adds a new section before it. Coordinators should specify section headings, not positional numbers, in template-modification ACs ⚠️ 1 source RF TFW-41/A §6 FC1 (executor observation) 2026-04-20
F19 review.md is the only TFW workflow that had a pre-existing Step 0 (Select Review Mode) before Session Naming was standardized. Any future editor inserting universal Step 0 content into review.md will face the same renumbering collision. Step 0 is now reserved for Session Naming; Select Review Mode = Step 1 by design ⚠️ 1 source RF TFW-41/B §6 FC1 (executor observation) 2026-04-20
F20 When HL deliverables list and TS AC checklist diverge (e.g., HL lists 15 terms, TS AC says 14), resolution = ask user. HL = vision (authoritative on WHAT), TS = spec (authoritative on HOW to verify). Neither auto-wins — user decides which is the actual contract ⚠️ 1 source RF TFW-41/D §6 FC1 (user resolution of ONB Q) 2026-04-20
F21 In multi-iteration research, all iteration dependencies are linear-sequential — iteration N reads only iteration N-1 RES. No DAG patterns observed across 8 iterations with 3 agents (AFD-2). Iteration number = dependency order. depends_on field in iterations.yaml is redundant ⚠️ 1 source RES TFW-42 iter1 FC2 (AFD-2 empirical data) 2026-04-30
F22 Generic capability guidance tables (e.g., "web research needs web search") are tautological overhead — removed by user as valueless. When agent field already provides traceability, adding a guidance matrix that doesn't name specific tools adds noise without helping coordinators decide. Over-engineering research findings into framework conventions = anti-pattern ⚠️ 1 source User decision, TFW-42 session (removed Agent selection guidance from conventions.md) 2026-04-30