Write an outcome-driven goal spec that drives agents toward verifiable outcomes — solid evidence, green-CI PRs, high code quality. Pulls context from the last 30 days of coding-CLI convos (/history) and cross-session memory (/ms) so the goal is grounded in real past failures, not invented acceptance criteria. Output is a .converge/goal.md ready to feed /h (goal_harness) or /a (finish-the-job). Grounded in a 30-day audit of 6,433 Hermes + 18,773 Codex sessions (2026-06-28).
Instrucciones de origen · Vista previa de solo lectura
name
write-goal
version
1.1.0
description
Write an outcome-driven goal spec that drives agents toward verifiable outcomes — solid evidence, green-CI PRs, high code quality. Pulls context from the last 30 days of coding-CLI convos (/history) and cross-session memory (/ms) so the goal is grounded in real past failures, not invented acceptance criteria. Output is a .converge/goal.md ready to feed /h (goal_harness) or /a (finish-the-job). Grounded in a 30-day audit of 6,433 Hermes + 18,773 Codex sessions (2026-06-28).
when_to_use
Use when the user types /write-goal <topic>, asks to 'write a goal spec for X', 'draft a goal for X', 'design a goal with evidence', or wants a goal doc that will actually drive agents to a green-PR outcome instead of stalling halfway.
[{"1.1.0 (2026-06-28)":"Actually ran /history + /ms against 6,433 Hermes sessions + 18,773 Codex threads + Claude Code JSONL. Added \"Ground truth\" section with real pattern counts (Skeptic QA 3,305 threads, Green CI 1,014 threads, /harness only 1 session in 30d). Added Phase 2.5 (/repro for bug-shaped goals). Added audit-grounded goal shapes (verify-completion, meta-goal, design/evaluate). Added 8 design decisions grounded in verbatim user goals from last month. Added mem0 memory findings."},{"1.0.0 (2026-06-28)":"Initial skill — mines last 30d of coding-CLI convos + memory, writes a structured outcome-driven goal spec wired for evidence/green-CI/quality. Output compatible with /goal_harness and /finish-the-job."}]
write-goal
Write a goal spec that drives agents to verifiable outcomes — solid evidence, green CI PRs, high code quality — grounded in your actual coding history, not invented acceptance criteria.
Why this skill exists
Three failure modes that keep recurring in the user's last month of coding-CLI sessions (May 22 → June 28, 2026):
Vague goal → vague outcome. Goal written as "fix the cold-start bug" produces an investigation summary, not a merged PR. No verifiable end-state is named up front, so the agent ships a write-up and stops.
No evidence contract. Goal says "make it work" but never says which evidence layer counts. Agent posts a green claim without verifying current-head CI green and no merge conflicts — gh pr checks alone can report stale passes or miss merge conflicts.
No project-specific quality bar. Goal omits the repo's own adversarial quality checks (ZFC, ZFC-leveling, root-cause-first, /es, /er). The agent thinks lint + unit tests are "done" and the user has to redo the work in a follow-up PR.
The fix: mine the last 30 days of coding-CLI convos + memory for the real failure shape and the real success shape on this codebase, then write a goal spec that encodes the specific evidence artifacts, CI gates, and quality checks the agent must produce. The goal becomes a contract the agent can drive itself against — /h, /a, /finish all consume it.
When to load this skill
User types /write-goal <topic>, or natural-language equivalent ("write a goal spec for X", "draft a goal that …")
User asks for a goal to be designed with explicit evidence + green-CI + quality requirements
User wants to take an ambiguous ask ("ship the new campaign wizard step") and turn it into a contract the next agent (or /h loop) can execute against
Pipeline (execute all phases in order; no optional skips)
Phase 1 — Intake (≤1 turn)
Collect the goal topic and any constraints. If the user gave the topic as $ARGUMENTS, use that verbatim. Otherwise ask ONE consolidate question:
What's the one outcome you want — the verifiable end-state (merged PR URL, verified local behavior, deployed service), not an activity?
Capture as inputs (write to a scratchpad, not the goal file yet):
topic — the noun at the center (a feature, a bug class, an infra change)
scope — repo(s) / file(s) / surface(s) in scope
outcome_shape — one of: green-pr-merged, pr-open-green, local-state-verified, dry-run-on-machine (these are the four provable end-states from finish-the-job — never invent a fifth)
hard_constraints — anything the user named (no API keys in code, ≤N tokens, must keep feature flag X off, etc.)
If the user typed /write-goal <topic> with no constraints at all → make the call yourself per the user's standing rule (finish-the-job § Anti-patterns: "correct but misinterpret is fine"), do not block on Phase 1 questions.
Phase 2 — Mine the last 30 days of coding-CLI history (mandatory)
Pull the user's actual coding history so the goal is grounded. Run two parallel searches:
A. History search (cross-CLI) — invoke /history <topic> --recent 30 --limit 20. Sources it covers (per ~/.claude/skills/history-search/SKILL.md):
Claude Code: ~/.claude/projects/*/*.jsonl
Codex: ~/.codex/state_5.sqlite threads
Hermes: ~/.hermes/state.db messages (FTS5)
What to extract from the history results:
Prior attempts on this topic — what was tried, what failed, what shape the failure took (test red, CI timeout, reviewer CHANGES_REQUESTED, agent stopped halfway, etc.)
The real success shape on this codebase — was "done" last time a merged PR + CI log, a screenshot + /er verdict, or a local run output? Capture the exact artifact type.
Pitfalls this repo has hit before — merge conflicts on the same files, flaky tests, quick "looks green" false positives, deny-list patterns from adoptive-external-agent-convention if the user has needed it.
The specific checks this repo runs — note any CI workflow names, advisory Green Gate logs, /es /er /code_standards invocation patterns that appear in past sessions.
B. Memory search — invoke /ms <topic> --recent 30 --source beads --limit 5 (and a second /ms <topic> --source hermes --limit 5 if beads is sparse). Pull:
User preferences (e.g. (h) no parallel PRs, (i) centralize repeated frontend strings as constants, (j) thread-roadmap cadence)
Tool quirks (gh pr checks can report stale passes—use statusCheckRollup + mergeable check for 2-gate /green, .beads/issues.jsonl is huge—use br, etc.)
Past failures with the same shape — these become the "Pitfalls" section in the goal.
If both history and memory return nothing → say so explicitly in the goal doc's "Prior Context" section: "No prior CLI sessions or memory entries matched <topic> in the last 30 days. Goal spec written from the user's input only; first run — pitfalls to be discovered during execution."
Phase 2.5 — Bug-shaped goals: run /repro first (if applicable)
From the 30-day audit: 319 of 6,433 Hermes sessions were test/repro sessions. When the goal is bug-shaped ("X isn't working", "Y is broken", "Z is failing"), add a "/repro step before Phase 3:
Load the repro skill (or repro-twin-clone-evidence skill for your-project.com)
Reproduce the bug FIRST — get the actual error/failure output — Before writing the goal doc
Put the repro output into the "Prior context" section of the goal
Skip this for: design goals, green-PR goals on existing branches, feature implementation goals, infra goals. Only bug-shaped goals need repro first.
Phase 3 — Select the outcome contract (one decision)
Pick the ONE end-state the agent must provably reach, drawing on Phase 2's history:
Additional goal shapes from the audit (map to the existing 4 states):
"Did X merge?" / "Is X finished?" → local-state-verified (verify with gh pr view / git log, not a new PR)
"Design/evaluate a plan for X" → dry-run-on-machine (the design doc IS the artifact; the user will execute if they approve)
"Fix the harness/skill behavior" → green-pr-merged if a code fix is needed, local-state-verified if only a config/skill file changed
"Why aren't these PRs green?" → pr-open-green (drive the existing PR to green, not a new PR)
Outcome shape
When to pick
Required proof artifact
green-pr-merged
Default for code changes that touch production paths
gh pr view <N> --json state = MERGED (verified at current-head SHA with CI green + no merge conflicts; Green Gate workflow logs are advisory only and are not aggregate /green proof) + Layer-2 evidence (env-preferences rule: unit-only is INSUFFICIENT for production changes)
pr-open-green
When the user must do the merge (e.g. protected branch, auth gate)
gh pr view <N> --json statusCheckRollup,mergeable shows mergeable == MERGEABLE + success-only, null-safe current-head statusCheckRollup[] (0 non-success checks excluding exact advisory Green Gate and Cursor Bugbot) + one-line blocker the user must clear
local-state-verified
For non-PR changes (config, ops, infra)
git diff --stat + git log --oneline -3 + actual command output captured in the final reply (described = fail; shown = pass)
dry-run-on-machine
For risky migrations / destructive changes the user will execute
Exact pasteable commands run against a fresh worktree, output captured, user can paste and get the same result
Never invent a fifth outcome shape. finish-the-job's contract has four; this skill stays inside that.
Phase 4 — Write the goal spec (the deliverable)
Write the goal to .converge/goal.md in the session's cwd (the path /goal_harness reads). Use the template at templates/goal-spec.md — copy it and fill in every field. All fields are mandatory; <TBD> is a fail. The template contents:
# Goal: <topic>## Outcome (verifiable, single end-state)<ONEof:green-pr-merged | pr-open-green | local-state-verified | dry-run-on-machine>### Proof artifact<exactartifacttheagentmustproduceatcompletion — e.g. "PR #Non $GITHUB_REPOSITORYmergedwithcurrent-headCIgreen + nomergeconflicts, /erPASSverdict (advisory), screenshotof /campaigns/newaftermerge">## Scope- Repo(s): <owner/repo>- Files / surfaces: <pathsorfeatureareas>- Out of scope: <whattheagentmustNOTtouch — protectsagainstscopecreep>## Hard constraints (non-negotiable)-<eachconstrainttheusernamedorthat'sencodedinmemory — e.g. "noparallelPRs; pushontoexistingPR'sheadbranchifoneexists", "centralizenewstringsinfrontend_constants.ts", "noANTHROPIC_API_KEYinai_universebackend", "usenvmNode22, neverHomebrewnode">## Evidence required- Layer: **Layer 2 integration minimum** (env-preferences rule — unit-only is INSUFFICIENT for production changes; <100 delta lines of non-test code is the only unit-only exception)
- Artifacts:
-<e.g. "realcallstacklogshowingthefixpathishit, mockonlyatexternalAPIboundaries"> -<e.g. "browserscreenshotoftheuser-facingsurfaceafterdeploy"> -<e.g. "/erverdictURLfromevidencereview">- Forbidden:
- Bullet list of command names only (proof-before-claim rule)
- "Key output observed: dependency install completed" (fabrication pattern)
- Green claim without verifying current-head CI green and no merge conflicts (env-preferences rule)
-`<60s` completion claim on a multi-step I/O task (proof-before-claim)
## Green-CI definition (specific to this repo)- Canonical /green 2-gate definition (current-head SHA):
- Gate 1: CI green — every CI check at HEAD passes (`statusCheckRollup` 0 non-success checks, excluding advisory `Green Gate` and `Cursor Bugbot`)
- Gate 2: No merge conflicts — `mergeable == MERGEABLE`- Green Gate workflow logs are advisory only and are not aggregate `/green` proof.
- CodeRabbit/Bugbot and review feedback remain optional advisory checks and do not block `/green`.
- Failure on CI or merge conflicts = iterate the fix, push to PR head branch (pr-ci-fix-autopush rule: push without being asked), re-check. Do NOT report "fixed" until the new head SHA shows both gates PASS.
## Quality bar (adversarial gates this repo uses)
Run all of these at completion (from /goal_harness):
- **/es** — Evidence Standards (both ~/.claude/skills/evidence-standards/SKILL.md AND .claude/skills/evidence-standards.md if the repo has one)
- **/er** — Evidence Review (adversarial, independent reviewer)
- **/code_standards** — 3 parallel lanes:
- ZFC (zero-framework cognition) — no keyword routing, no regex intent detection, no hand-tuned scoring
- ZFC-leveling — no level-up fields leaking into non-leveled surfaces
- root-cause-first — prompt/schema fixes tried before backend protection; document why if backend enforcement is added
- **Independent Agent Review** — full-diff code review by an isolated subagent
- 4/4 PASS = DONE. Any FAIL = fix & loop. Same score 2 iterations = STALLED → escalate to human.
## Dispatch routing
- Inline if: single tool call OR tight sequence with no fork, <10 lines changed
- AO worker (`ao spawn` via dispatch-task) if: multi-step code work, PR creation, worktree isolation needed, CI iteration loop
- `/h` (goal_harness) if: user wants the full 4-gate adversarial loop to wave the goal in
-`/a` (finish-the-job) if: user wants hands-off single-task drive to the end-state
- Do NOT self-execute anything that crosses the write/PR threshold inline (SOUL.md scope-pivot-to-ao commit)
## Definition of Done
At the end, the agent's final reply MUST contain (finish-the-job Phase 4):
1. End-state declaration: "✅ Done: <greenPR #Nmerged> | <PR #Nopen + green, awaitingyourreview> | <localstateXverified> | <dry-runoutputcaptured>"
2. Proof artifact: PR URL as markdown hyperlink `[#<N>](https://github.com/<owner>/<repo>/pull/<N>)` — never bare `#N` (pr-hyperlink rule)
3. Mid-stream judgment calls: every decision the agent made instead of asking, with one-line rationale
4. NO follow-up question — "want me to X?" is a violation
## Pitfalls (mined from history — Phase 2)-<pitfall1 — e.g. "PRmergeabilityflippedtoCONFLICTINGaftermainbranchupdated — re-verifymergeablestatusatcurrentHEAD">-<pitfall2 — e.g. "mergeconflictsonworld_logic.pyrecurbecauseit'sa421KBfile — alwayspull--rebasebeforepush">-<pitfall3 — e.g. "UnittestspassbutLayer2integrationfails — env-preferencessaysunit-onlyisINSUFFICIENTforproductionchanges">## Prior context (Phase 2 output summary)- History searches: <count> sessions matched "<topic>" in last 30 days
- Memory searches: <count> matching entries (beads + hermes)
- Key findings:
-<finding1 — e.g. "PR #7832hitthesamefileinMay23session; samefixshapeapplies"> -<finding2>- If no prior context: state explicitly "No prior CLI sessions or memory matched; first run on this topic."
## Hand-off- Goal file written to: `.converge/goal.md` (this file)
- To execute: `/h` (run /goal_harness against this goal) — or `/a` for hands-off single-task drive — or `ao spawn` via dispatch-task for raw worker dispatch
- This goal spec is consumable by finish-the-job, goal_harness, and dark-factory — all three read `.converge/goal.md`
Phase 5 — Quality scaffold the spec (mandatory self-check)
Before declaring the goal written, run these gates on your own draft:
Outcome verifiability — is the "Outcome" line ONE of the four named end-states, with a concrete artifact? If it says "improve" / "address" / "investigate" / "explore" — rewrite it. Those are activities, not outcomes.
Evidence specificity — does "Evidence required" name the actual artifact (file:line, PR URL, screenshot path, log excerpt), not a category ("tests", "logs")?
Green-CI grounded — does "Green-CI definition" list this repo's actual workflow names from Phase 2? If Phase 2 found none, write "first run — gates to be enumerated from gh run list --workflow on first execution" instead of inventing gate names.
Pitfalls sourced — does every "Pitfalls" entry cite a real Phase 2 finding (date, session id, or memory key)? Invented pitfalls are fabrication — strike them.
Definition of Done includes the no-follow-up rule — verbatim: "NO follow-up question" appears in the DoD section.
Dispatch routing references the real commits — does it cite scope-pivot-to-ao / pr-green-dispatch so a worker reading the goal knows when to escalate inline→AO?
If any gate fails → rewrite the goal section in question. Do not post a half-spec.
Phase 6 — Hand off (final reply)
Post a single reply containing:
Goal spec path — ~/.converge/goal.md (or repo-local .converge/goal.md) as an absolute path with file:// scheme (env-preferences rule)
One-line outcome summary — what the agent drives toward
Provenance — how many history + memory entries fed the spec, dated
Hand-off command — one of:
/h — to run the full adversarial 4-gate loop (/goal_harness)
ao spawn --task "$(cat .converge/goal.md)" — raw worker dispatch
~/.hermes/scripts/gh-safe-publish issue create --title "<topic>" --body-file .converge/goal.md — issue-track the goal for babysitting
What was decided mid-stream — any judgment call you made (which outcome shape, which gates to include, which pitfalls to surface)
NO follow-up question — finish-the-job's final-reply contract applies here too
Pitfalls (do NOT do these)
❌ Inventing acceptance criteria — gates, workflow names, pitfall shape that don't appear in Phase 2 history are fabrication. If Phase 2 returned nothing for a field, write "first run — to be discovered during execution" rather than guessing.
❌ Activity-shaped outcomes — "improve cold start", "investigate the BQ gap", "address reviewer feedback" are activities. Rewrite as "PR #N merged with cold-start p50 < Xs measured by / CI green + no merge conflicts / /er PASS (advisory)".
❌ Unit-only evidence for production changes — env-preferences: unit-only is INSUFFICIENT for production changes (Layer 2 integration minimum). Only the <100 delta lines of non-test code exception applies.
❌ gh pr checks as green proof — env-preferences: gh pr checks can report stale passes; must verify current-head statusCheckRollup (CI green) and mergeable == MERGEABLE (no conflicts).
❌ Goal with no dispatch routing — without an inline-vs-AO rule, a future agent will self-execute a 30-commit PR inline and drop it when the gateway session caps. Cite scope-pivot-to-ao.
❌ Goal with no "no follow-up question" in DoD — finish-the-job's contract requires the agent drive to conclusion. A goal that allows "want me to X?" mid-stream reproduces the silent-stop pattern.
❌ Skipping Phase 2 mining — without history+memory grounding the goal is generic; the third failure mode (vague goal → vague outcome) recurs. Specific failure mode (2026-06-28): substituting session_search (a different tool that searches message summaries) for actual /history queries against the SQLite databases. session_search does NOT query state.db or state_5.sqlite — it has different scope and returns different results. Phase 2 MUST use session-history-search to query the actual session stores with the SQL patterns documented in that skill.
❌ Writing project-specific workflow names into the skill itself — the SKILL.md stays general; the workflow names live in the generated .converge/goal.md. The skill is reusable across repos.
❌ Bare #N PR references — pr-hyperlink rule: every PR number in the goal doc itself and in the final reply must be a markdown hyperlink.
Verification (the skill works when…)
✅ Generated .converge/goal.md has all template fields filled (no <TBD>)
✅ Outcome line is one of the four named end-states with a concrete proof artifact
✅ At least one Pitfalls entry cites a real Phase 2 finding (session id, date, or memory key)
✅ Green-CI definition either lists real workflow names from history OR explicitly says "first run"
✅ DoD section contains "NO follow-up question" verbatim
✅ Final reply paths use the file:// absolute-path scheme for the goal doc
✅ A future session invoking /h or /a against the goal produces a green-PR-merged (or other end-state) outcome without needing a mid-stream clarifying question
evidence-standards (Phase 4 evidence-requirements section — load to know what counts as evidence)
pr-green-definition (Phase 4 green-CI section — canonical 2-gate PR green definition: CI green + no merge conflicts)
ponytail (Phase 4 quality bar — referenced by /code_standards)
root-cause-first (Phase 4 quality bar — referenced by /code_standards)
harness (Phase 6 hand-off — if user picks /h)
Ground truth — 30-day history audit (2026-06-28)
Ran /history + /ms against all three session stores on 2026-06-28 before designing this skill. These are the real findings that shaped the skill design.
Real user goals from the last month (verbatim from session first-messages)
These are the actual goal-shaped messages Jeffrey sent (filtered from cron/test/ping noise):
"Make or find the PR where we make all local servers default to agy cli" — outcome: PR found/created, verified merged
"did this PR truly merge and is agy_provider.py on origin main?" — verification goal, needs proof not activity
"fullrun lets confirm agy cli is the default cli" — autonomous drive request
"Why arent these PRs /green?" — frustration with CI blocking, needs 2-gate green fix
"Run /repro — other people in the battle aren't taking their turns" — bug repro + fix
"Lets design/evaluate a plan to convert game state json into a game state ledger" — design/evaluate goal, not a PR
"Review this evidence using /er and /advice — how do we increase parallelism?" — evidence review + meta question
"Look at the test and CI failures and merge conflicts across all PRs last 48 hours" — triage/babysit goal
"cleanup merge conflicts and /green the PR" — drive-to-green goal
"Dont do the task do the meta task. Lets read this thread and see why hermes misunderstood /green and run /harness" — meta-goal: fix the harness behavior, not just the bug
"Is this finished? Lets do 1-4" — verification + drive forward
Memory findings (mem0 search)
"Can drive agents toward outcomes with solid evidence, and values high code quality and green CI PRs" (score 0.82)
"Can identify anti-patterns in agent behavior" (score 0.71)
"Likes designing and evaluating plans" (score 0.65)
"Can make drive decisions autonomously" (score 0.66)
"Prefers to present options and let the user respond" — BUT finish-the-job overrides this for execution
Design decisions grounded in the audit
Goal shapes must include "verify completion" — goal #2 ("did this PR merge?") and #11 ("Is this finished?") are verification goals, not code goals. The existing 4-state contract handles this via local-state-verified but the skill should explicitly call it out.
Meta-goals are real — goal #10 ("Dont do the task do the meta task") is a harness-fix, not a code-fix. The skill must handle meta-goal shapes (fix the harness/skill/agent behavior, not the code).
Evidence review is the dominant adversarial pattern (3,305 Codex threads). The skill must wire evidence deeply — not just "run tests" but route to the actual evidence review (/er) pipeline the user already uses.
Green verification is a top-3 activity (1,014 Codex threads). The skill must encode the canonical 2-gate green procedure (current-head CI green via statusCheckRollup + no merge conflicts via mergeable, with Green Gate logs and review feedback treated as advisory).
/harness is severely underused (1 session in 30 days). This skill should make it easy to produce goal specs that feed /h, increasing harness adoption.
Repro/testing patterns are common (319 sessions). The skill should know when a goal needs /repro as a Phase-1 step (bug-shaped goals, not feature-shaped).
Design/evaluate goals are valid — goal #6 ("design/evaluate a plan") doesn't produce a PR. The skill must support dry-run-on-machine as a valid outcome for architecture/design goals.
The user's actual phrasing is casual and outcome-oriented — not formal ("fix the thing", "did this merge?", "why aren't these green"). The skill must accept informal phrasing and translate to the structured contract.
Worked example — 2026-06-28 (this session)
User invoked /write-goal after a month of recurring CI/deploy issues that never shipped.
Phase 1 intake: topic = "fix the auto-deploy precompute OOM", outcome_shape = green-pr-merged, scope = $GITHUB_REPOSITORY (embed pipeline), hard_constraints = "no parallel PRs; if PR #7880 is open, push onto its head branch".
Phase 2 mining:/history "precompute OOM" returned 3 Slack sessions + 1 CLI session; /ms "embed batch size" surfaced the recipe doc at ~/.hermes/skills/worldarchitect/references/auto-deploy-dev-precompute-oom-investigation.md and ~/.hermes/skills/devops/production-vs-main-drift/references/precompute-oom-deploy-workflow-2026-06-23.md. Findings: prior sessions diagnosed + wrote a fix recipe but never pushed to main; deploy has been failing 80+ days.
Phase 3: outcome_shape = green-pr-merged (the recipe is a destructive change to _embed_texts; user wants it on main, not a worktree dry-run).
Phase 4: goal doc names the exact EMBED_BATCH_SIZE change, the deploy.sh exit-137 detection, the current-head CI requirements, and the prior-failure pitfalls (recipe written twice, never pushed).
Phase 5: quality scaffold passes — outcome is verifiable, evidence is Layer 2 (real callstack on embed path), gates named from history, pitfalls cite session ids.
Phase 6: reply posts file://.../.converge/goal.md + hands off /h precompute OOM fix or ao spawn --task "$(cat .converge/goal.md)".
The goal became the input for the /harness (or /a) loop that finally drove the fix to a merged PR — closing the 80-day failure the user identified in SOUL.md COMMIT: diagnosis-with-fix-not-shipped.
