Philosophy of Trace and Trace-First Workflow

When work disappears

AI can produce a plan, analysis, decision, design, or program in minutes. Yet the visible result is often the smallest part of the work. Before it appears, someone identifies what matters, frames the problem, tests assumptions, rejects plausible alternatives, makes trade-offs, and decides what “good” means. If only the output survives, that intellectual work becomes invisible.

The loss appears at the next handoff. A new person or agent sees the result but not the purpose it served, the limits that shaped it, or the uncertainty that remains. They must find the original specialist, reconstruct an old conversation, or repeat earlier mistakes. Fast generation then creates more isolated outputs without creating continuity.

This is the problem TFW addresses: when people delegate part of intellectual work to AI agents, how can the work remain understandable, governable, inspectable, and continuable?

Chat is raw material, not memory

A chat history is useful. It contains questions, corrections, candidate decisions, and context that may exist nowhere else. But chronology is not authority. An early answer may have been superseded; a confident claim may be unverified; the accepted result may be buried among abandoned branches.

Organizational memory requires selection. It must distinguish what matters now from what merely happened, and it must remain available to the people and systems entitled to use it. A service conversation can contribute to that memory, but no vendor’s chat should be its only home.

What a Trace is

A Trace is a selected, durable, inspectable material record of work. It preserves enough of the purpose, material context, boundaries, decisions, result or current state, and continuation point for another authorized person or agent to inspect the work and proceed.

A Trace is not a transcript of everything said. It does not require hidden chain-of-thought, pretend to contain complete memory, or turn a recorded statement into truth. It may be a short note, a connected set of documents, a decision record, evidence, or another form suited to the work. Its value is functional: it makes the work easier to reconstruct, compare, correct, hand over, and continue.

A good Trace does not promise that repeating a process will produce an identical result. Conditions, models, and judgment change. It preserves the grounds needed to understand what was done, evaluate the current result, and make the next decision without replaying the entire original conversation.

What makes work continuable

Continuity begins before execution. A Goal names the change sought. Value says for whom and why that change matters. A Question exposes an uncertainty instead of hiding it behind a premature answer. A Task bounds a piece of work. Context supplies the material needed to understand it, while Boundaries say what the executor may decide, what is out of scope, who may accept the result, and when to stop or escalate.

These ideas are not a demand for one universal form. They are a completeness test for delegation. The person keeps purpose and legitimate authority. The agent receives bounded work. The work leaves a selected Trace that includes its result or current state. It ends with an authoritative next step or an explicit close.

Evidence, review, and verified knowledge add assurance when consequences justify their cost. A small, reversible task may need a compact trace and a direct check. High-risk work may require independent review, stronger evidence, and stricter authority boundaries. More ceremony is not automatically more reliable.

What remains human

Delegation transfers work, not responsibility. People or legitimate institutions retain purpose, authority, judgment, acceptance, accountability, and the responsibility to stop. An agent can analyze, propose, create, test, and coordinate inside the authority it has been given; it cannot create legitimate authority for itself.

Useful delegation does not require a person to master every domain detail in advance. It requires enough understanding to state the purpose, provide material context, set boundaries and acceptance conditions, identify the accepting authority, and define when uncertainty or risk must return to a human.

Observability makes correction possible; it does not guarantee correctness. Likewise, ownership and access to work memory depend on context. Employees, employers, clients, data subjects, and other parties may have legitimate privacy, security, contractual, or intellectual-property interests. The stable rule is narrower: important memory must not be trapped by default inside the service that helped produce it.

Work that can describe its own state

A TFW project is sometimes called self-aware in an operational sense. From durable traces, it can answer:

1. What are we trying to do and why?
2. What do we currently know, assume, and not know?
3. Which material decisions were made, rejected, or superseded, and on what evidence?
4. What is the current state of the work?
5. Who or what has authority for the next decision?
6. How can another human or agent continue without reconstructing the original chat?

This capability implies neither consciousness nor independent authority or accountability. It means the state of work is inspectable. This repository is self-applying because it uses its own traces to change its method, and self-describing because those traces explain the change. It is a reference example, not the only valid realization.

What TFW is

The Philosophy of Trace is the foundation: significant delegated intellectual work should leave selected continuity, because an isolated output is not enough to govern, inspect, or continue the work.

Trace-First Workflow is a methodology for joint work between people and AI when people delegate part of intellectual work to agents. People retain purpose, legitimate authority, judgment, acceptance, accountability, and stop responsibility; agents perform bounded work; selected durable traces preserve enough material context, decisions, result or current state, and continuation for the work to be inspected and resumed.

Workflows and implementations realize that methodology under particular conditions. The word methodology is an owner-selected position expressed through this function, composition, and its exclusions; it is not a claim that research proved a new category or that TFW invented provenance, documentation, oversight, or knowledge management.

Proportional realizations

Light, Assisted, and Full preserve the same semantic continuity at different levels of coordination and assurance. Light keeps the minimum visible by hand. Assisted adds bounded support where routine discipline is likely to fail. Full adds explicit roles, gates, evidence, and review for work whose complexity or risk warrants them.

Moving from one visible problem to the next has worked as one founder-led teaching route. It is not a universal maturity law or proof of superior learning, retention, transfer, adoption, or field effect. A realization should be chosen for the work and risk, not as a badge of advancement.

NS1 — Purpose

TFW exists to protect purposeful, human-governed continuity when intellectual work is delegated to AI. Its value is not the production of more text or more process. Its value is that another authorized person or agent can understand what the work is for, inspect its material grounds and current result, see where authority remains, and continue without rebuilding the original conversation.

Work that increases output while obscuring purpose, authority, inspectability, or continuation is contrary to this North Star even if it is polished and technically correct.

NS2 — Principles

1. Purpose before activity. State the Goal and Value before optimizing Tasks; reject work that cannot explain the change it serves.
2. The Saint-Exupéry Principle. “Perfection is achieved not when there is nothing left to add, but when there is nothing left to remove.” Seek the simplest complete and coherent form. Subtraction is not improvement when it damages purpose, value, correctness, architecture, modularity, inspectability, or continuation.
3. Questions before premature answers. Make consequential uncertainty visible; do not let fluent output conceal an undefined problem.
4. Selected Trace, not transcript. Preserve material context, decisions, state, and continuation; do not archive everything or demand private reasoning.
5. Human authority, bounded delegation. Name boundaries, acceptance authority, accountability, stop conditions, and escalation before granting autonomy.
6. Continuation over isolated output. Finish with an inspectable result or current state and an authoritative next step or explicit close.
7. Assurance proportional to risk. Add evidence, review, and durable verified knowledge when their expected value exceeds their cost; subtract ceremony that does not protect the purpose.

Each principle must change a work choice. If it only decorates a document, it is not functioning as a principle.

Methodology values

Four values govern how these principles are practiced:

• Candor Over Flattery. Disagree when evidence supports it, surface risk and uncertainty, and never substitute sycophantic agreement for judgment.
• Structural Enforcement. Put important gates in observable structure or state, not in prose exhortation alone; a rule that cannot reveal its own violation is only advice.
• Naming Creates Behavior. Use precise terms that cue the intended cognitive role and action, while remembering that a name can guide behavior but cannot guarantee compliance.
• Portability. Keep durable project context in ordinary, provider-independent files or equivalent inspectable forms; no vendor may become the sole home of project memory.

Operational routes

Standard /tfw-economics connects intended and achieved product work to observed AI resources: executive explanation, sourced product map, inspectable numbers, grounded improvements and coverage. Its canonical workflow provides complete prose/table reports; visuals are supplementary and PDF is explicit-request-only outside the repository. Daily supplies cumulative own-source snapshots before each orderly return and a final economic summary, preserving its ordinary local form and authority. Economics follows default pinned manifest distribution; Daily discovery remains explicitly optional. Both preserve the existing formal roles and acceptance routes.

Use /tfw-update for a pinned, receiver-safe upstream update. Each attempt leaves an immutable receipt under .tfw/update_receipts/; the receipt records what was applied, preserved, skipped, or refused and never replaces task-local state.

Use /tfw-release only when the project has a release contract or an explicitly selected release effect. RELEASE.md is optional project context, not a universal TFW versioning or publication rule; the workflow records the selected effect and its evidence before any external action.

The selected-Trace principle preserves material intent, rationale, and continuity without treating code or any other output as disposable or identically regenerable. A complete, usable, bounded result plus an explicit close or continuation carries completeness over speed. Visible uncertainty, source boundaries, and refusal to fabricate evidence carry honesty over convincingness. Where truth belongs assigns one authoritative owner to each kind of truth rather than forcing every truth into one monolithic file.

NS3 — Non-goals

TFW must not become:

• a prompt collection, raw chat archive, or store of every intermediate message;
• a deterministic generator that treats code or other outputs as disposable and promises identical reproduction;
• a replacement for human purpose, judgment, legitimate authority, acceptance, accountability, or stop decisions;
• a maximum-documentation bureaucracy or documentation factory that measures success by artifact count;
• a vendor-bound tool, runtime, model, interface, or memory feature;
• a vehicle for untested claims of novelty, correctness, comprehension, learning, autonomy, adoption, or other capability.

Success Criteria

A TFW project is succeeding when:

1. An authorized participant can resume from a durable checkpoint. A person or agent can inspect the relevant state and continue without reconstructing the original chat; this does not promise lossless context.
2. Material decisions are traceable. The project preserves consequential choices, grounds, alternatives, and dispositions rather than claiming that every message or minor action deserves a permanent record.
3. Verified knowledge can compound. Reviewed findings can improve later decisions and onboarding; raw notes and fluent claims do not become truth automatically.
4. The result is ready for acceptance. It is complete, usable, inspectable, free of placeholders, and does not require the recipient to reconstruct missing work, while authorized human judgment and acceptance remain necessary.

Where truth belongs

This Project North Star owns TFW’s stable purpose, principles, boundaries, and non-goals. Current reference-implementation mechanics belong to the conventions, glossary, workflows, templates, and Editions. Start with the Quick Start when the practical question is how to use TFW.

Selected task-local traces, verified knowledge, the changelog, and Git history preserve why the project became what it is. A future approved and versioned Body of Knowledge may expand concepts, contradictions, hypotheses, and evidence limits, but it will remain subordinate to this North Star and requires a separate task. A working draft outside this repository is not current authority.

That separation lets stable meaning remain clear while implementations, evidence, and explanations continue to change.

Standard /tfw-economics is installed/updated by default with selected adapters through the manifest and core workflow. It connects results to unique resources, uses sourced conditional API prices and preserves task-owned used quote/report epochs. Public provider/month observations live at ~/.tfw/rates, fresh for the exact model/profile for less than 30 days; lookup uses working first-party tools. Daily remains optional; four formal roles and existing authority/acceptance/source routes remain. Old optional/custom installations follow the core migration guide before connected writes. Neither API equivalents nor cache freshness prove subscription charges, historical unchanged prices, ROI or native cross-host behavior.