decide
Skill0 starsChanged 53 days agoMITView on GitHub
Structured decision framework: commit a consequential NEW choice -- investment, career, life, or architecture -- to a permanent decision-log record. Use when weighing an offer, a capital deploy or exit, or a build commitment, or when a pre-mortem, reversibility check, or real decision record is wanted. Distinct from /challenge -- /decide commits a NEW choice and writes the record; /challenge only stress-tests an EXISTING thesis and writes none. Not a lookup of past decisions.
53 other projects use a version of this.
Adopted by
Most recently taken up by
- fullstackhouse/skillstheir own version, first seen Nov 15, 2026
- warrendeanlangeveldt/code-kittheir own version, first seen Nov 15, 2026
- Integralist/agent-skillstheir own version, first seen Nov 14, 2026
- aditya-vermax/skillstheir own version, first seen Nov 8, 2026
- vimkim/my-skillstheir own version, first seen Nov 8, 2026
- yuri-xyz/skillstheir own version, first seen Nov 8, 2026
Used with
Projects with it also use redesign, weekly and 8 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
arjunkshahSkillredesign
Redesign existing interfaces to premium quality while preserving product goals and functional behavior.
Also in 3 projects that use this
gaiaaiagentSkillweekly
Generate a weekly digest rolling up dailies into higher-level pattern analysis
Also in 3 projects that use this
SethGammonSkillexperiment
Automated optimization loop with scalar fitness function. Proposes changes in isolated worktrees, measures with a metric command, keeps improvements, discards failures. Supports convergence detection and diminishing returns.
Also in 3 projects that use this
kv0906Skillinbox
Quick capture mode (with content) or batch processing mode (no content). GTD-style routing of inbox items to proper folders. Use "/inbox thought to capture" or "/inbox" to process.
Also in 3 projects that use this
genomewalkerSkillhealth
Soul system health check with remediation
Also in 3 projects that use this
AWSSkillmap
Map verified service capabilities against customer-provided control framework and threat model. Produces mapping-results.json, controls-catalog.md, and framework-mapping.md. Supports pluggable frameworks: CCMv4, NIST 800-53, CIS Benchmarks, ISO 27001, etc. Split into 4 focused sub-skills for better accuracy; this skill orchestrates them.
Also in 3 projects that use this
Other versions53 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 53
- ParadevAI/Paradev★ 16ksame name
- superset-sh/superset★ 16ksame name
- coreyhaines31/makerskills★ 916same name
- me2resh/apexyard★ 542same name
- menkesu/awesome-pm-skills★ 441same name
- BlockRunAI/blockrun-mcp★ 390same name
- kv0906/pm-kit★ 138same name
- Abilityai/cornelius★ 109same name
- me2resh/agent-decision-record★ 53same name
- alirezarezvani/gaios★ 50same name
43 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository18 other files it gives its agents.
The file
Show its rules and text
We split this skill into the rules it gives an agent
23 rules. 0 are common, 10 are distinctive, and 13 are very specific to this project.
0
Common
Most of the 52 edited versions we read keep these.
10
Distinctive
This author's own choices. Worth a look.
13
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 31
If information is missing, say what's missing and whether the decision should wait.
- Distinctive: line 32
Check decision-log.md for related prior decisions.
- Distinctive: line 33
Always identify the "smallest possible bet" that tests the thesis with minimal risk.
- Distinctive: line 40
Use: investment / career / life / skill-architecture decisions warranting permanent record + adversarial pre-mortem.
- Distinctive: line 42
Not for: trivial choices (use prose discussion).
- Very specific: line 79
If both calls fail or return []: emit Phase 0 -- no vault context retrieved and continue.
- Very specific: line 90
U1 residency gate (orchestration economics, AGENTS.md): Run python tools/run-share.py --current-context FIRST.
- Very specific: line 134
Use the Agent tool with subagent_type vault-researcher.
- Very specific: line 156
Use the Agent tool with subagent_type decision-critic.
- Very specific: line 170
Use the Agent tool with subagent_type thesis-critic.
- Distinctive: line 172
If decision is NOT investment-shaped (career, life, health, vault-architecture): skip this dispatch -- domain-mismatch is documented skip, …
- Very specific: line 227
Vault-resolved (in ALL_BASENAMES): keep wikilink (legitimate vault reference)
- Very specific: line 229
Placeholder syntax (
<X>,..., bareX, etc.): leave unchanged (vault-audit.py filters) - Distinctive: line 232
After stripping, post-validation invariant: each composed section body must contain ZERO unresolved wikilinks.
- Distinctive: line 274
If decision affects pending items (creates new pending item, removes existing pending item, changes patient-cash deploy strategy, etc.):
- Very specific: line 275
Update Pending Items section additive-only (removal only on explicit completed: X markers per /retro convention).
- Very specific: line 276
Update Active Context if decision changes session focus.
- Very specific: line 277
Update frontmatter updated: field to today.
- Distinctive: line 279
If decision is purely advisory (no pending-item impact): skip hot.md update entirely.
- Distinctive: line 283
After Phase E-J compose all 4 target file modifications IN MEMORY but BEFORE Phase K atomic write:
- Very specific: line 285
Write each composed file to a tmp dir (e.g., wiki/research/test-tmp/.precheck/decide-<slug>/).
- Very specific: line 319
Write Calendar/decisions/decision-<slug>-<date>[-HHMM].md (NEW detail file)
- Very specific: line 354
NEVER write to: .raw/, private/, finance/, credentials/.
“Common” is measured: kept by at least 6 in 10 of the 52 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Execution Rules
- Subagent dispatch is MANDATORY when specified by phase AND the running harness exposes an Agent tool (Phase D.0 dispatches
vault-researcherfor prior-decision baseline; Phase E.0 dispatchesdecision-criticfor adversarial pre-mortem; Phase E.5 dispatchesthesis-criticCONDITIONALLY when decision is investment-shaped). Pre-emptive skip based on judgment ("vault context not needed", "I can do my own pre-mortem", "this isn't really investment-shaped") is FORBIDDEN. The subagents decide applicability via their own N/A return contracts. Legitimate fallback fires ONLY on (a) contract violation -- subagent return missing required fields (calibrated %, smallest-possible-bet, reversibility verdict, ITT score for thesis-critic), (b) actual dispatch failure -- timeout, rate limit, tool-denial, hard subagent crash, OR (c) NO Agent tool in the running harness -- the documented inline fallback runs verbatim and is reported asinline-no-Agent-tool, which is NOT a DEVIATION. Pre-emptive skip (Agent tool present, dispatch not attempted) surfaces in the phase audit as DEVIATION. Quality preserved by additive design: existing inline pre-mortem + reversibility logic at Phase E.3-E.4 remains intact as legitimate-failure fallback. - Subagent legs (vault-researcher, decision-critic, thesis-critic) carry their own model pins in their agent definitions wherever an Agent tool exists. No phase of this skill branches on model identity.
Discipline and quality
Quality Rules
- Take a position. The output is a recommendation, not a menu.
- The pre-mortem is the most valuable section. Make it genuinely adversarial; not polite.
- Dollar amounts from current-session broker read evidence, never estimated.
- Calibrated confidence: percentages with justification, not HIGH/MEDIUM/LOW alone.
- If information is missing, say what's missing and whether the decision should wait.
- Check decision-log.md for related prior decisions; avoid contradicting yourself without acknowledging the change.
- Always identify the "smallest possible bet" that tests the thesis with minimal risk.
- F11 Phase C discipline mandatory: set
.claude/state/auto-commit-disabledBEFORE any Write/Edit; clear ONLY after atomic commit succeeds. - Body-preservation sha256 on every append-mode UPDATE (decision-log, daily, hot.md). Content outside insertion sites byte-exact post-write.
- Pattern 22 ASCII-only on new content (em-dash -> --, smart quotes -> straight, ellipsis -> ...).
When to use / not
Use: investment / career / life / skill-architecture decisions warranting permanent record + adversarial pre-mortem.
Not for: trivial choices (use prose discussion); thesis stress-tests (use /challenge); routine status-checks (use /brief); session retrospectives (use /retro).
Invocation
Mode routing (Pattern 6 deterministic; invocation modes)
| Syntax | Behavior |
/decide <decision> | Full framework: Phase 0 retrieval -> Phases A-E analysis (vault-researcher + decision-critic dispatch where an Agent tool exists) -> Phases F-J compose -> Phase J.0 gate -> Phases M-N atomic 4-file write + commit |
/decide <decision> --preview | dry-run; render decision text + planned writes to stdout; SKIP all writes; F11 never set |
/decide <decision> --confirm | pause before each Write for user confirmation |
/decide <decision> --no-decision-log | skip decision-log table append (e.g., for purely-exploratory pre-mortems) |
/decide <decision> --replace | explicit same-day overwrite of an existing detail record (default on collision is the -HHMM variant per Phase A.5) |
Process (15 atomic operations)
Phase 0 -- Vault context retrieval (Phase 3.6d -- runs BEFORE Phase A; read-only; runs in --preview)
Skill-level semantic retrieval over the local HNSW vault index (covers wiki/ + Calendar/ + private/; Atlas/ is NOT indexed, so Phase D's inline domain refs are still read directly -- this block COMPLEMENTS, not replaces). Deeper and path-filtered vs the universal UserPromptSubmit hook. Read-only: a Bash node call pair; writes nothing; safe in --preview. Complements Phase D.0's vault-researcher dispatch (semantic-vector recall vs the subagent's targeted grep). Let D = the decision argument.
0.1 -- Build 12 decision-decomposition queries:
"<D>", "<D> options and alternatives", "<D> risks and failure modes", "<D> similar prior decisions", "<D> reversibility", "<D> opportunity cost", "<D> counter-arguments", "<D> time horizon", "<D> calibration base rates", "<D> invalidation signals", "<D> downstream dependencies", "<D> stakeholder impact".
0.2 -- Fire query-skill.mjs TWICE via Bash, piping the JSON payload to stdin (heredoc; no temp file). Model loads once per call (~0.3-1.5s):
- BROAD (cross-vault, no filter): stdin
{"queries":[<the 12>],"top_k":100,"threshold":0.60}tonode ~\.vault-substrate\query-skill.mjs - FOCUSED (decision record subtree): stdin
{"queries":[<the 12>],"top_k":25,"threshold":0.60,"filter_path_prefix":"Calendar/decisions/"}to the same script.
Each returns [{path,line,score,text}]; the script always exits 0 (emits [] on error) so Phase 0 never crashes the run.
0.3 -- Merge BROAD + FOCUSED; dedup by path:line (keep higher score); sort score desc; cap top 100. Store as VAULT_CONTEXT.
0.4 -- EMIT a visible summary (MANDATORY):
Phase 0 -- retrieved vault context (N hits; M from Calendar/decisions/)
Top results:
[score] path:line -- first ~80 chars of text (up to 15)If both calls fail or return []: emit Phase 0 -- no vault context retrieved and continue. NEVER HALT on Phase 0.
0.5 -- Downstream consumers (phases UNCHANGED; consult VAULT_CONTEXT instead of re-deriving):
- Phase D.0/D.1 (Context loading): seed prior-decision baseline + doctrine refs;
vault-researcherfocuses deltas. - Phase E.1/E.2 (First principles + Options): surface prior similar decisions + outcomes as base rates.
- Phase E.3 (Pre-Mortem): seed failure-mode precedents from prior decision post-mortems in VAULT_CONTEXT.
Cite any used chunk inline as (vault: path:line) -- plain text, NO wikilink syntax (Phase 0 context is raw, not vault-resolved).
Phase A -- Pre-flight state verification
U1 residency gate (orchestration economics, AGENTS.md). Run
python tools/run-share.py --current-context FIRST. Exit 1 means the session is
carrying more resident context than this run should re-pay on every turn: report
the number, recommend /clear + re-invoke in a fresh session, and proceed only if
the operator says to. Cost basis: cache_read is ~82% of frontier cost and equals
the sum over turns of the resident prefix, so a heavy run started at 500K residency
costs multiples of the same run started fresh. This is a RECOMMENDATION, not a HALT
-- never block work the operator asked for; make the price visible before it is paid.
- Parse
<decision>argument. - Resolve today's ISO date.
- Compute slug from decision title (lowercase-kebab, <=40 chars).
- Resolve OUTPUT_PATH =
Calendar/decisions/decision-<slug>-<date>.md. - Same-day collision check: if exists, default
-HHMMvariant;--replacereserved. - F11 collision check (
.claude/state/auto-commit-disabledexists -> HALT with diagnostic). - Pre-edit sha256 capture for append-mode files: decision-log.md, today's daily note, wiki/hot.md.
Phase B -- State-transition print (BEFORE F11)
Emit planned reads + planned writes + collision-handled output paths to stdout. Format:
/decide -- planned state transition:
reads: current-session broker read evidence / decision-log.md (prior context); domain-specific
refs (macro-outlook, watchlist, portfolio-doctrine for investment; etc.); hot.md (pending items)
writes:
- Calendar/decisions/decision-`<slug>`-`<date>`[-HHMM].md (NEW)
- Calendar/decisions/decision-log.md (append; then `python tools/gen-ledger-views.py --ledger decisions --ingest` per OSANWE-V2 ADR-02)
- Calendar/daily/`<today>`.md (append to ## Decisions section)
- wiki/hot.md (frontmatter updated; ## Pending Items conditional update)
F11 set after this printPhase C -- F11 flag set (MANDATORY before any Write/Edit)
touch .claude/state/auto-commit-disabled. This blocks the PostToolUse auto-commit hook so the 4 file mutations land as a single atomic commit at Phase L.
Phase D -- Context loading (parallel reads)
Two legs. D.0 delegates the prior-decision baseline to vault-researcher where an Agent tool exists; D.1 always runs the domain-specific inline reads, narrowed to deltas when D.0 returned a valid synthesis and widened to cover the baseline when it did not.
Phase D.0: DELEGATED dispatch to vault-researcher (PREFERRED prior-decision baseline; Phase C wiring 2026-05-02)
MANDATORY when the harness exposes an Agent tool (per Execution Rules): dispatch first, no pre-emptive skip. The subagent pre-loads prior similar decisions + sessions-log context + relevant doctrine, freeing inline reads to focus on deltas.
Use the Agent tool with subagent_type vault-researcher. Pass input: {query: "<decision-argument-keywords>", scope: "Calendar/decisions/ + Atlas/concepts/ + wiki/hot.md", session_id: "<current-session-id>"}. Expected return: ranked citation table (file + line + excerpt + relevance HIGH/MED/LOW) + 3-5 sentence synthesis covering prior similar decisions and their outcomes, active context from hot.md, and applicable doctrine references.
Validate return: citation table present + synthesis 3-5 sentences. On contract violation OR dispatch failure: fall through to inline reads below. Surface in the phase audit.
No Agent tool in the running harness: skip the dispatch and run the Phase D.1 inline reads below as the documented fallback, widening them to cover the prior-decision baseline the subagent would have supplied (Calendar/decisions/decision-log.md + sessions-log.md + wiki/hot.md, seeded by VAULT_CONTEXT from Phase 0). Report mode: inline-no-Agent-tool -- this is NOT a DEVIATION.
On dispatch success: use synthesis to inform inline reads (focus on deltas + domain-specific files only).
Phase D.1: Inline domain-specific reads (always runs; informed by D.0 synthesis if available)
Domain detection from decision argument. Load relevant files in parallel:
- Investment decisions: current-session broker read evidence, Atlas/concepts/investing/macro-outlook.md, Atlas/concepts/investing/watchlist.md, Atlas/sources/investing/ref-portfolio-doctrine.md, relevant entity notes from wiki/entities/tickers/
- Any decision: Calendar/decisions/decision-log.md (last 30 days; check for related prior), wiki/hot.md (pending items + active context), Calendar/decisions/sessions-log.md (last 7 days; recent ratifications)
Phase E -- Decision analysis (5-section spine)
Order of operations: the dispatch legs (E.0 pre-mortem, E.5 conditional investment stress-test) and the E.6 GATE-F prerequisite resolve FIRST; the five composed sections E.1-E.5 are then written using whatever those legs returned (or their documented inline fallbacks).
Phase E.0: DELEGATED dispatch to decision-critic (PREFERRED pre-mortem path; Phase C wiring 2026-05-02)
MANDATORY when the harness exposes an Agent tool (per Execution Rules): dispatch first to populate Phase E.3 Pre-Mortem section, no pre-emptive skip. decision-critic generates 3-5 calibrated failure modes + smallest-possible-bet + reversibility verdict adversarially.
Use the Agent tool with subagent_type decision-critic. Pass input: {decision_argument: "<from arg>", horizon: "<6mo default or per-domain>", reversibility_hint: "<from initial framing>"}. Expected return: ranked failure modes (calibrated % + detectability + recoverability + cascade) + smallest-possible-bet (cost in $ + days + decision-rule) + reversibility verdict (REVERSIBLE | NEAR-IRREVERSIBLE | ONE-WAY).
Validate return: calibrated % on every failure mode, smallest-bet has explicit decision rule, reversibility uppercase per contract.
On contract violation OR dispatch failure: fall through to inline E.3 Pre-Mortem + E.4 Reversibility composition below.
No Agent tool in the running harness: skip the dispatch and compose E.3 Pre-Mortem + E.4 Reversibility inline verbatim as documented below, holding the same contract the subagent would have returned (calibrated % on every failure mode, smallest-possible-bet with an explicit decision rule, uppercase reversibility verdict). Report mode: inline-no-Agent-tool -- this is NOT a DEVIATION.
On dispatch success: use returned failure modes for E.3, smallest-bet integrates into E.4-E.5, reversibility verdict drives Phase J hot.md update.
Phase E.5: DELEGATED dispatch to thesis-critic (CONDITIONAL; when decision is investment-shaped)
MANDATORY when applicable AND the harness exposes an Agent tool (per Execution Rules): if the decision is investment-shaped (large position add, thesis-driven sell, doctrine-ceiling bet, tax-advantaged cash deployment), dispatch thesis-critic for asset-thesis stress-test. Decision-critic covers the general pre-mortem; thesis-critic adds bear-evidence-first ITT discipline specific to investment theses.
Use the Agent tool with subagent_type thesis-critic. Pass input: {ticker: "<extracted from decision>", primary_thesis: "<thesis-slug>", primary_drivers: [<from decision-argument>], market_regime: "<inferred or from prior briefing>", proposed_rating: "<from decision direction>"}. Expected return: 3-5 ranked failure modes with ITT self-score >=7/10.
If decision is NOT investment-shaped (career, life, health, vault-architecture): skip this dispatch -- domain-mismatch is documented skip, not deviation.
No Agent tool in the running harness: on an investment-shaped decision, run the bear-evidence-first lens inline inside E.3 -- load invalidation evidence BEFORE confirming evidence, check the ref-portfolio-doctrine invalidation triggers and the thesis triggers: block per-row, and self-score the Ideological Turing Test (>=7/10; below 7 -> strengthen the bear case, never ship). Report mode: inline-no-Agent-tool -- this is NOT a DEVIATION.
Phase E.6: GATE-F prerequisite (judgment-gates kit, 2026-07-06; position decisions only)
If the decision proposes an ADD / TRIM / EXIT on a specific instrument: a same-day GATE-F sheet (verdict != BLOCKED) is a PREREQUISITE to ratification. Run the GATE-F procedure per the /gate skill's ref-gate-tables.md (fill markers with provenance, python tools/gate-eval.py --compute, write sheet, --check, registry row) and cite the sheet in the detail record's Related. BLOCKED -> the decision records DEFER with the gate mandates as review triggers (a legitimate close), never a ratified action. FOMO-SUSPECT -> the recommendation carries the cooling-off-48h + tranche-cap mandates verbatim. Non-position decisions: skip (documented skip, not deviation). Enforced by HALT item 16a.
E.1 First Principles Decomposition
- What exactly is being decided? (specific framing, not vague)
- What are the constraints? (money, time, information, reversibility)
- What assumptions are embedded in the framing?
- What would have to be true for each option to be the right choice?
E.2 Options Analysis
For each viable option (target 2-4 options, not menu of 10):
- Expected outcome (base case): specific dollar / outcome statement
- Best case scenario + probability (with justification)
- Worst case scenario + probability (with justification)
- Opportunity cost: what you give up by choosing this
E.3 Pre-Mortem (CRITICAL section; 3-5 failure modes for leading option)
Frame: "It is 6 months later. This decision was catastrophic -- the worst credible outcome. Write the post-mortem as if explaining to an investor what happened."
For EACH failure mode:
- Probability (X%) with specific reasoning
- Was this detectable beforehand? What signal was ignored or unavailable?
- Recoverable or permanent?
- Dollar/career/life cost (specific numbers from actual portfolio data or salary projections)
- Cascade: what else fails if this fails? (correlated positions, dependent decisions)
E.4 Reversibility Assessment
- Reversible? How easily? At what cost?
- Cost of waiting for more information vs acting now?
- "Smallest possible bet" version that tests thesis?
E.5 Recommendation
Single clear recommendation with:
- Action (specific, not vague)
- Confidence: X% -- [evidence basis]. Would change to Y% if [condition].
- Single strongest reason to act
- Single strongest reason to wait
- Smallest possible bet
- Review trigger: specific measurable event that would change recommendation
- Review date: YYYY-MM-DD when to revisit regardless
Phase F -- Body-scope wikilink validation (v2.0; /retro v2.2 parity)
For each section of the composed decision text (E.1 through E.5), scan body for [[<target>]] patterns:
- Vault-resolved (in ALL_BASENAMES): keep wikilink (legitimate vault reference)
- MEMORY_PREFIXES match (
memory:,feedback_,auto_,user_,project_,reference_,research_): rewrite as[[memory:<stem>]]grammar (visually distinct, machine-detectable, exempt from broken-wikilink checks) - Placeholder syntax (
<X>,..., bareX, etc.): leave unchanged (vault-audit.py filters) - Else: strip wikilink brackets, leaving plain text reference. Log strip count to per-decision provenance.
After stripping, post-validation invariant: each composed section body must contain ZERO unresolved wikilinks. This prevents the C2-C5 class of regressions (decision-log accumulating broken wikilinks from conversation context).
Phase G -- Compose decision-log table append
Append a single row to Calendar/decisions/decision-log.md:
| `<date>` | `<domain>` | <decision summary> | <rationale 1-sentence> | `<status>` |Per-file body-preservation gate: pre-edit sha256(decision-log.md) recorded at Phase A.7. Post-edit sha256(content-before-insertion-window) MUST equal pre-edit sha256 of same window. HALT and rollback if differs.
Phase H -- Compose decision-<slug>-<date>.md detailed record
Frontmatter:
---
categories: [decisions]
type: decision
created: `<today>`
updated: `<today>`
status: active
confidence: <0-100>
tags:
- topic/`<domain>`
- decision/`<slug>`
related:
- "decision-log"
- "hot"
- "`<domain>`-moc" -- e.g., investing-moc
---Body sections: H1 + Decision header (date, domain, conviction) + The Question + Options table + Pre-Mortem + Reversibility + Recommendation + Review Triggers + Related Prior Decisions (wikilinks to past decision-{slug} files where applicable).
Phase I -- Compose daily note append (## Decisions section)
Append a single row to today's Calendar/daily/<today>.md ## Decisions table:
| `<domain>` | <decision summary> | `<why>` |Per-file body-preservation gate: same sha256 invariant.
Phase J -- Compose hot.md update (CONDITIONAL)
If decision affects pending items (creates new pending item, removes existing pending item, changes patient-cash deploy strategy, etc.):
- Update
Pending Itemssection additive-only (removal only on explicitcompleted: Xmarkers per /retro convention). - Update
Active Contextif decision changes session focus. - Update frontmatter
updated:field to today.
If decision is purely advisory (no pending-item impact): skip hot.md update entirely.
Phase J.0 -- Pre-commit /vault audit gate (v2.0 prevention architecture)
After Phase E-J compose all 4 target file modifications IN MEMORY but BEFORE Phase K atomic write:
- Write each composed file to a tmp dir (e.g.,
wiki/research/test-tmp/.precheck/decide-<slug>/). - Run
python tools/skill-precheck.py <tmp1> <tmp2> <tmp3> <tmp4> --skill /decide. - Parse exit code. If 0: proceed to Phase K. If 2: HALT with diagnostic.
- Auto-repair pathway: if HALT was due to broken wikilinks, re-run Phase F body-scope validation with stricter mechanical de-link, then re-attempt gate. If second attempt fails, surface diagnostic and halt for user input.
- Bypass:
CLAUDE_VAULT_BYPASS_VALIDATOR=1env var (logged to.claude/state/bypasses-<date>.log).
This is defense-in-depth on top of PreToolUse pre-write-validator.py + PostToolUse wikilink-check.py / frontmatter-check.py / orphan-check.py. The Phase J.0 gate prevents broken composition from reaching disk in the first place.
Phase K -- Pre-Output HALT gate (16-item invariant check)
- F11 set
- OUTPUT_PATH collision resolved (
-HHMMif needed) - Frontmatter complete and schema-compliant on detail file
- Decision-log row composed
- Detail file body composed with all 6 sections (Question / Options / Pre-Mortem / Reversibility / Recommendation / Review Triggers)
- Pre-mortem has 3-5 failure modes; each with probability + detectability + recoverability + cascade
- Recommendation has confidence percentage + condition for change + smallest-possible-bet
- Body-scope wikilink validation passed (Phase F applied; zero unresolved wikilinks)
- Phase J.0 pre-commit /vault audit gate passed (skill-precheck.py exit 0)
- Sha256 body-preservation invariant computed for each append-mode file
- ASCII-only on new content (Pattern 22)
related:field has 3-5 wikilinks (decision-log, hot, domain MOC, related prior decisions)- Tag vocabulary guardrail: tags use
topic/*,decision/*; no forbiddendomain/*ortype/*namespaces - Path-guard: no Write targets
.raw/,private/,finance/,credentials/ - Sessions-log entry composed for /retro pickup (the decision is itself a substantive item)
- F17 Co-Authored-By absent in planned commit body 16a. GATE-F prerequisite (Phase E.6): position decisions carry a same-day GATE-F sheet with verdict != BLOCKED, cited in Related; FOMO-SUSPECT mandates appear verbatim in the recommendation. Non-position decisions: N/A.
Phase L -- ASCII pre-write scan (Pattern 22)
Byte-scan all NEW content. HALT on byte >127. Replace em-dash (-> --), smart quotes (-> straight), ellipsis (-> ...).
Phase M -- Atomic write (4 files)
- Write
Calendar/decisions/decision-<slug>-<date>[-HHMM].md(NEW detail file) - Edit
Calendar/decisions/decision-log.md(append table row; sha256 verify) - Edit
Calendar/daily/<today>.md(append ## Decisions row; sha256 verify) - Edit
wiki/hot.md(conditional update per Phase J; sha256 verify)
F.halt: if any Edit fails (e.g., old_string anchor changed since Phase A pre-flight), IMMEDIATE HALT, F11 stays on, structured report (succeeded / failed / not-attempted), no partial commit. User can git checkout -- <orphan> to rollback or fix root cause and re-invoke.
Phase N -- Atomic commit (F14 narrow stage + F17 verify)
- F14 narrow stage:
git addexactly the 4 touched files (nogit add -A). Commit with message:
agent(decide): `<slug>` -- `<domain>` -- `<conviction>` <2-3 sentence rationale summary> Smallest bet: `<description>` Review trigger: `<event>` Review date: `<YYYY-MM-DD>`- F17 verify: post-commit
git log -1 --format=%B | grep -c Co-Authored-ByMUST be 0.
Phase O -- F11 clear
rm .claude/state/auto-commit-disabled. PostToolUse auto-commit hook re-enabled for subsequent skills.
Idempotency (v2 spec requirement)
- Decision marker signature:
<slug>-<date>-- same slug + date = idempotent re-run - Pre-persist check Phase G: if proposed decision-log row's signature already present, SKIP (log "already captured")
- Detail file: same-day collision ->
-HHMMvariant per Phase A.5 - Idempotent re-run produces zero diff
Path-guards (mechanical)
NEVER write to: .raw/, private/, finance/, credentials/. Never delete (move-to-archive only via the archive directory). All writes route through Calendar/decisions/, Calendar/daily/<today>.md (append-only), wiki/hot.md (frontmatter + Pending Items conditional update).
Tag vocabulary guardrail
Decision file tags: topic/<domain> (e.g., topic/investing, topic/career), decision/<slug>. Forbidden: domain/*, type/*. Frontmatter MUST have categories: [decisions] and type: decision.
Coordination with other skills
- /brief: reads decision-log.md last 30 days for "Today's Focus" priors. /decide outputs feed /brief next-morning context.
- /retro: sessions-log entry composed at Phase K item 15 surfaces /decide invocations as substantive ratifications. /retro extracts and persists.
- /invest: portfolio-related /decide invocations should reference /invest entity notes in
related:. /invest may itself trigger /decide for "should I add this position" questions. - /challenge: /decide produces a recommendation; /challenge stress-tests it. The pair: /challenge surfaces invalidations -> /decide ratifies the response.
Failure taxonomy
| Failure | Phase | Symptom | Recovery |
| F11 collision | A | another skill in-flight | Wait OR rm .claude/state/auto-commit-disabled if confirmed orphan |
| Same-day collision | A | OUTPUT_PATH exists | Default -HHMM; user can --replace |
| No Agent tool in harness | D.0 / E.0 / E.5 | harness exposes no dispatch primitive | Run the documented inline fallback verbatim; report inline-no-Agent-tool (NOT a DEVIATION) |
| Phase F unresolved wikilink | F | broken wikilink in composed text after de-link | HALT; report target + line; user fix or auto-strip |
| Phase J.0 GATE finding | J.0 | skill-precheck.py exit 2 | HALT; report findings; auto-repair retry; user input on second fail |
| sha256 body-preservation violation | M | content outside insertion window changed | F.halt; back out the Edit; sha256-preserving retry |
| Mid-batch Write failure | M | first Write succeeded, second failed | F.halt; F11 stays on; structured report; user git checkout -- orphan |
| ASCII byte >127 | L | non-ASCII in NEW content | Replace and retry; do NOT proceed |
| F17 verify fail | N | Co-Authored-By in commit body | git commit --amend to remove |
Examples
Example 1: Investment decision (tax-advantaged-account cash deploy)
/decide Should I deploy $<amount> retirement-account cash now? ->
- Phase D loads current-session broker read evidence, macro-outlook.md, watchlist.md, portfolio-doctrine.md
- Phase E.3 pre-mortem 5 failure modes (macro-shock delay, thesis concentration drift, valuation compression, etc.) with %, detectability, cascade
- Phase E.5 recommendation: e.g., HOLD the cash and stage limit orders at pre-set levels
- Phases G-J atomic 4-file write
- Sessions-log entry composed for /retro
Example 2: Architecture decision (build vs adopt)
/decide Build a local retrieval index or adopt a hosted one? ->
- Phase D loads the retrieval design notes, the current index benchmarks and the relevant ADRs
- Phase E.4 reversibility: adopting a hosted index moves data off the machine; building locks in maintenance
- Phase E.5 smallest bet: benchmark both on one fixed query set before committing
Example 3: --preview mode
/decide Add <TICKER> position --preview -> renders full analysis to stdout; no Write; no commit; user reviews and either invokes without --preview to commit or iterates.
Example 4: Empty-decision safeguard
/decide with no argument -> HALT with diagnostic "decision argument required; usage: /decide '<question>'".
Example 5: Idempotent re-run
/decide Should I deploy retirement-account cash now? invoked twice in same session ->
- Second run Phase G dedup via marker signature
should-deploy-retirement-account-cash-now-<date>matches existing row -> SKIP decision-log append - Detail file collision ->
-HHMMvariant (or--replaceif user explicitly wants to overwrite)
Related skills
/retro-- session-end retrospective; consumes /decide ratifications/brief-- reads decision-log.md priors; /decide outputs feed next briefing/invest-- portfolio-decision pair; /invest analysis -> /decide ratification/challenge-- thesis stress-test; /challenge surfaces invalidations -> /decide ratifies response/vault-- audit consumer; /decide outputs validated against GATE classifiers
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

