Goal-Driven

A mini methodology for initiatives where the destination is fixed but the path isn't. Human owns the compass; agent walks the path and keeps the log.

GOAL.md is the institutional memory of why. The record is the institutional memory of what was tried. Together they let a fresh agent — or future-you — pick up a months-long initiative without losing direction.

Commands

When invoked with an argument, dispatch to the corresponding file:

• /goal-driven set → Read and follow commands/set.md. First run: interview the human and produce the initial GOAL.md, plus scaffolding. Re-run: update mode — refreshes the agent-config blocks (including cross-references to sibling skills installed since), re-checks plumbing. Doesn't re-interview the goal; for re-assessment use review.
• /goal-driven review → Read and follow commands/review.md. Strategic checkpoint plus protocol maintenance: re-assess where the project stands against GOAL, surface drift, propose corrections.
• /goal-driven close → Read and follow commands/close.md. Wrap up a finished initiative: draft retrospective, mark GOAL closed, archive goals/.
• No argument → continue with the methodology below (the normal loop).

Which command when:

• New initiative → set
• A sibling skill was installed/removed and agent-config cross- references feel stale → set again (update mode; no re-interview)
• ≥ 2 weeks since last review, record feels overgrown, midpoint of an explicit timeline, or you sense you've drifted → review
• Initiative is done, abandoned, or superseded → close
• Mid-work, just logging progress → no argument (normal loop)

When to use this skill (and when not)

Good fit:

• Multi-week or longer; ≥ 3 uncertain decision points likely
• The path is genuinely unclear — you'd rewrite a detailed plan in 2 weeks
• Success criteria can be stated, even if some are observable proxies

Bad fit:

• One-off scripts, bug fixes, week-long features with a clear plan → use a TODO list
• Pure execution where the path is known and only the doing remains → use design-driven blueprints
• Goals you can't articulate at all → spend an hour writing first; come back

For personal use specifically. This skill works for personal initiatives that look like projects, not for general life management. The threshold is four conditions held together: multi-week+ horizon; articulable falsifiable criteria (or paired proxies); willingness to engage with the agent ≥ weekly; the cost of failure warrants the friction. Below that — vague aspirations ("be healthier", "write more"), weekly habits, generic quarterly planning — a notebook or habit tracker serves better. Concrete fits: book / article series, learning projects with a shippable output, job search, side-business launch, training plans with concrete metrics. Concrete misfits: anything where "how would I know I'm done?" can't be answered without hand-waving.

Directory structure

project/
└── goals/
    ├── GOAL.md                  ← Stable compass (General Line, criteria, invariants)
    ├── OPEN-STOPS.md            ← Index of unresolved STOP signals
    ├── record-2026-04.md       ← Past month
    └── record-2026-05.md       ← Current month (auto-rotates monthly)

Two stable files (GOAL.md, OPEN-STOPS.md) plus monthly records. No archive/ directory by default — old months stay in place. If goals/ becomes crowded after a year, move them under goals/record-archive/ manually; the convention is informal.

The compass / path asymmetry

GOAL.md is the compass. It changes rarely, by deliberate human decision. The path is everything else: what you're trying this week, what you tried last week, what worked, what didn't.

The path mutates constantly. The compass mutates only when:

• A success criterion turns out to be the wrong measure (replace it)
• The General Line itself is questioned by new evidence (rethink it)

Both are deliberate, human-approved events. They are NOT what happens when "I tried X and it didn't work, let me try Y" — that's just walking.

This asymmetry is the whole point. If GOAL.md and the record change at the same rate, you don't have a compass; you have a notebook.

Permission gradient

File	Who writes	When	Human's role
GOAL.md	agent	Only at initial set or explicit GOAL change	Approves each section, line by line
record-YYYY-MM.md	agent	End of session, or on STOP / rotation	Reviews entry draft in chat before write
OPEN-STOPS.md	agent	When a STOP is created or resolved	Confirms the line in chat
STOP signals	agent surfaces	When trigger condition hit	Decides: change path / change goal / agent misjudged

Never let the agent silently edit GOAL.md. Every change must be echoed in chat first. Even adding a single criterion is a deliberate event.

The agent is not a passive scribe — it drafts, proposes, surfaces. The human is not a writer — they approve, redirect, decide. This asymmetry prevents two failure modes at once: human laziness ("I'll write the goal later") and agent drift (silent rewording into something subtly different).

The three-moment protocol

The agent's default behavior — drafting, exploring, summarizing — is usually fine. But three moments bias toward silent drift if left to defaults: when first articulating the goal (premature commitment to fuzzy wording), when ending a work session (forgetting what just happened and why it matters to the criteria), and when evidence threatens the goal (rationalizing past inconvenient signals to keep moving). The protocol intervenes at exactly those moments. Outside them, normal work.

Moment 1 — Setting the goal (creating GOAL.md)

The agent interviews the human via chat. It does not draft GOAL.md alone and present it for review. It asks, in order:

1. General Line — "In one or two sentences, what should be true when this is done?"
2. Success criteria — for each thing implied: "How would you know that's achieved?" Push for falsifiable; if soft, push for a proxy indicator.
3. Invariants — "What must stay true regardless of how we get there?"
4. Non-goals — "What's tempting but explicitly out of scope? What would an outsider lobby for that we've already decided against, and why?" Push back if a candidate is just an inverted goal — non-goals are forbidden zones, not goals restated in the negative.

The agent echoes each section back as it's drafted, gets confirmation, moves on. GOAL.md is written only after all sections are confirmed. See commands/set.md for the full interview script.

Moment 2 — End of every work session

Before the session ends, the agent drafts a record entry in the chat:

Entry draft:
- What I did: ...
- Observations: ...
- Criteria check:
    C1 ✓ (cite specific evidence)
    C2 ✗ (cite specific evidence)
    C3 unclear (no observation this session)
- Judgment: path-level / goal-level / no change
Confirm and append?

The human confirms or edits in chat. Only then does the agent append to the current month's record.

The criteria check is the heart of the discipline. Each ✓/✗ must cite a concrete observation from the session. C1 ✓ without evidence is forbidden — it must be C1 ✓ (ran 30 queries, recall 76%) or C1 unclear (no test this session). Bare verdicts hide drift.

Verdicts include time. Some criteria are met or not at the moment they're checked (latency under 500ms either is or isn't true today). Others accumulate toward a deadline ("200 paid subscribers by Sep 30"), and in the run-up they're neither met nor failed — they're on a trajectory. Judge them by trajectory: one plainly on pace is ✓ (on pace: <evidence>); one whose trajectory has visibly diverged is ✗ (off pace: <observation> projects <gap>). The evidence rule still applies — "on pace" needs an observation, not optimism. This is the same verdict vocabulary, used with awareness that what counts as "served" depends on when the criterion is supposed to land.

Identify the principal tension. Attention is finite. At any moment one tension dominates — the one whose resolution would most unblock the rest. Name it in the judgment ("this week the binding constraint is C1 pace; C2 stable; C3 not under pressure") rather than treating all criteria symmetrically; this tells the next session where to spend attention first. The dominant tension shifts as the project moves through phases — re-identify it each session from current observations, not from habit.

If a session is interrupted before the human confirms. Append the draft directly with a [unconfirmed draft] prefix in the title; the next session ratifies or revises it before continuing other work. A protocol that breaks when sessions don't close cleanly is worse than one that absorbs the common case.

Moment 3 — STOP signals

Two types of STOP, each with a different escalation path. Both halt work until the human decides.

Type A — Criterion not served. A criterion is failing now, or evidence makes its future failure visible in time to act. The point of STOP is to surface infeasibility while there's still room to redirect — waiting for a deadline to confirm what trajectory already shows wastes that room.

• Example (current violation): "C2: P95 < 500ms" but production shows 720ms after three attempts at different storage layers.
• Example (predictable violation): "C1: ≥ 200 paid subs by Sep 30" with 22 at end of month one — linear extrapolation lands near 66, not 200. The criterion isn't violated yet, but the path is.
• Agent surfaces in chat, proposes three options:
  • (a) Change path (try Y instead of X)
  • (b) Change criterion (require human-approved GOAL edit)
  • (c) Agent misjudged

Type B — General Line questioned. Criteria are met (or close), but new evidence suggests the goal itself solves the wrong problem.

• Example: criteria all green, but user research reveals users wanted topic clustering, not semantic search.
• Agent surfaces in chat, proposes:
  • (a) Reframe the General Line (deliberate GOAL edit)
  • (b) Stay course because the evidence is weak

Critical: STOPs are never silently logged and walked past. The agent must surface them in chat AND wait for the human's choice before continuing work. A STOP entry in the record without a chat exchange is a protocol violation.

Anti-flattery: the evidence rule

Agents writing their own progress reports have an obvious bias toward ✓ verdicts (they want to keep working, not stop and wait for review). The defenses, in order of strength:

1. Every ✓ must cite an observation from THIS session. Not "criterion still reasonable", not "no reason to think it broke". A specific thing that happened.
2. If no observation touches a criterion this session, the verdict is unclear, not ✓. Default to unclear. ✓ is earned, not assumed.
3. unclear accumulating across many sessions is itself a signal — that criterion isn't being measured. The review command catches this and nominates such criteria for retirement or for explicit instrumentation.

This rule is the single biggest reason this skill might survive contact with reality. Without it, the criteria check decays into checkbox theater within weeks.

Maintenance

Three things rot if untended: record volume grows unboundedly, open STOPs across files get forgotten, GOAL.md drifts inconsistent after edits. The protocol handles each — month-bounded record files cap volume once the project is long enough to warrant them, OPEN-STOPS.md indexes unresolved STOPs across files when more than one is in flight, and /goal-driven review periodically reconciles all three plus re-assesses strategy. See commands/review.md for what review covers; references/templates.md for record, STOP, OPEN-STOPS, and carry-over formats.

Review when ≥ 2 weeks pass since the last one, ≥ 30 new entries accumulate, the project hits the midpoint of its explicit timeline, a fresh agent picks up the project, or things just feel off.

Stories — opt-in interpretation layer

GOAL.md is normative and terse — good for fast reading and for keeping the compass hard to change accidentally, but insufficient when a criterion needs context, a choice needs justification, or an ambiguous term needs unpacking. The optional goals/stories/ directory is the second layer: per-topic narrative documents that interpret GOAL.md's terse text.

Stories are opt-in. /goal-driven set does not create the stories/ directory; the first approved story brings it into being. For initiatives with self-explanatory criteria, skip the layer — it's overhead. The agent proposes a story only when there's something nontrivial to interpret.

For trigger conditions, modification protocol, distinction from record, and cross-reference patterns, see references/stories.md. The per-file template (filename patterns, body shape) lives in references/templates.md.

Standalone vs paired with design-driven

Goal-driven works alone. Shape decisions, when they come up in an exploratory project, live as observations or named alternatives in the record — no separate proposal flow needed. This is the default mode and what the rest of this document assumes.

If design-driven is also installed in the project, the two divide labor: goal-driven owns why and how-far; design-driven owns what-shape. Each manages its own files; cross-references go by ID, not content (a record entry says "adopted decision 003"; a decision says "blocks goal STOP 2026-05-02 if rejected").

When to bring in design-driven from a goal-only project. The signal is: you can describe the system's shape in 2–3 sentences and an outside engineer would build the same mental model. At that point, run /design-driven bootstrap to extract the implicit shape into DESIGN.md. Goal-driven keeps running alongside; design-driven takes over the "how" question. Until that signal, defer — premature DESIGN.md gets rewritten.

When to bring in reframe instead of design-driven. If the initiative is in unsettled paradigm territory (AI-native, agent-first, post-mobile) and "what shape?" is ill-formed because no precedent exists, use /reframe before design-driven. Reframe redraws the skeleton from the new paradigm's primitives in concepts/<target>.md; once stress and comprehension tests pass, /reframe close graduates the skeleton into DESIGN.md and the design-driven flow takes over. In the unsettled phase, GOAL.md and concepts/<target>.md run in parallel — destination compass and shape working-theory.

Four interaction points to watch when both are present:

1. Goal pivot crosses design boundaries → also open a design/decisions/NNN-*.md proposal. The record entry doesn't replace the design proposal flow.
2. Design proposal would violate GOAL invariants → trigger a goal Type A STOP first; don't let design adopt and then quietly fail criteria.
3. STOP resolution requires shape change → resolve the goal STOP first, then open a design decision.
4. Cross-checks during periodic inspection — goal-driven's review surfaces design decisions not reflected in the record; design-driven's audit surfaces GOAL invariants potentially at risk. (The two skills use different command names — review vs audit — for the same periodic-inspection role.)

A third sibling skill — evidence-driven — overlays build-time discipline on top of design-driven (TDD cycle, falsifiable verification, evidence- trail State). Goal-driven doesn't read or write to evidence-driven's domain directly, but the rigor it provides at the execution layer makes the upward feedback loop trustworthy: STOP signals citing build observations are only credible if those observations have evidence behind them.

Structure follows need

The scaffolding around GOAL.md — OPEN-STOPS.md, monthly record rotation, periodic reviews — earns its cost when it's load-bearing. A short, single-stream initiative with no STOPs in sight doesn't need an OPEN-STOPS index; the one STOP, if it ever appears, lives visibly in the record. A six-week project doesn't need monthly rotation; a single record.md carries it until it doesn't.

Add structure when the absence starts to hurt: create OPEN-STOPS.md when a second open STOP exists and you can no longer hold both in mind; rotate to monthly files when one record becomes too long to scan. The directory layout shown above is the mature shape, not the starting shape — start small and grow into it.

GOAL.md is the exception. It's the compass, not scaffolding. Even a two-week initiative benefits from articulating it. If GOAL.md feels like overhead for what you're doing, this skill isn't what you need; use a notebook.

Example walkthrough

For a concrete end-to-end example — set, several record entries, two STOP scenarios, resolution — see references/example.md.