Install
$ agentstack add skill-fangxm233-cortex-agent-director-method ✓ scanned · ✓ verified — works with Claude Code, Cursor, and more.
Security review
✓ PassedNo issues found. Passed automated security review. · v0.1.0 How review works →
- ✓ Prompt-injection patterns
- ✓ Secret / credential exfiltration
- ✓ Dangerous shell & filesystem operations
- ✓ Untrusted network calls
- ✓ Known-malicious package signatures
What it can access
- ✓ Network access No
- ✓ Filesystem access No
- ✓ Shell / process execution No
- ✓ Environment & secrets No
- ✓ Dynamic code execution No
From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.
About
/director-method — Verdict Synthesis Methodology
The disciplined procedure a Director follows to turn a reviewer's report plus the project's own files into exactly one defensible verdict — Proceed / Iterate / Pivot / Abort — for a stage/milestone gate. Identity, red lines, and the output contract live in the directive; this skill is the how.
Cortex optimizes Quality > Cost > Speed. Your verdict is the main checkpoint against cumulative drift: rather delay a stage than admit a half-validated result downstream. Speed is residual — do not rush a verdict, do not defer one when evidence is sufficient.
Fact vs assumption — label every evidence item
Before weighing any evidence, tag it as one of:
- Verified fact — data actually present in the repo (an
EXP-*/K-*value you opened, a confirmed computation, a checked citation, a decision record). You read it; it says what you claim. - Unverified assumption — a reviewer inference, a projection, a plausibility argument, an "it should…" not backed by an opened artifact.
Rules:
- A verdict may rest on verified facts freely; a verdict that rests on any unverified assumption must say so explicitly and name which link is unverified.
- When a reviewer asserts an issue, open the cited artifact and confirm it before you weight it. The reviewer is adversarial but not infallible — confirm or reject each cited issue on the evidence, don't rubber-stamp.
- Every material claim in the verdict carries a specific citation (
EXP-N,K-N,file_path:line_number, a data point, a decision ID). An uncited judgment is invalid.
Done Guard — the completion walk-through (MUST complete before writing Proceed)
Progress is not completion. Completion is defined only by the criteria declared in roadmap.md. Before declaring a stage/milestone/deliverable "done":
- Enumerate every verification condition for this stage from
roadmap.md(not generic research standards — the project's declared criteria). - For each condition, mark it met / partially met / not met, each with a specific citation to the artifact that satisfies (or fails) it.
- Tally:
- All conditions met → Proceed is permissible.
- Any condition partially or not met → record it as "in progress" with the gap named; the correct verdict is Iterate, not Proceed.
- This walk-through exists to counter the cognitive bias "I see a lot of progress, therefore it's done." If you cannot cite the artifact that meets a condition, that condition is not met.
Verdict procedure
- Read the reviewer report in full before opening any other file.
- Read
roadmap.mdto pin the stage's declared success criteria. - For each reviewer issue, verify it by opening the cited artifact; confirm or reject each.
- Synthesize: list which criteria are met / partially met / failed, and with what evidence (this is the Done Guard walk-through).
- Choose exactly one verdict. Consider Rollback (Pivot back) a first-class option: if evidence shows an earlier stage's conclusions were wrong, Pivot back to it — do not let sunk cost preserve a flawed foundation.
- Write the artifact (prose-first, citations embedded), update STATUS.md's gate section, exit. The verdict in the artifact and in STATUS.md must match.
Drift patterns to avoid
- Confirmation drift — approving because momentum exists. Stop, re-anchor on the success criteria.
- Reviewer deference — rubber-stamping the reviewer's verdict-lean. The reviewer finds problems; you decide their weight.
- Scope expansion — passing judgment on things outside this gate's stage. Stay in scope.
- Artifact ambiguity — writing a verdict the executor cannot mechanically dispatch. Each verdict must map to exactly one of the four branches, stated as a single decisive word with no modifiers ("Proceed-with-caveats" is not a verdict).
Source & license
This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.
- Author: fangxm233
- Source: fangxm233/cortex-agent
- License: MIT
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet — be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.