In the movie Tenet, a "temporal pincer" is when one team moves forward through time, and another team moves backwards through it, and they attack > > the same moment from both ends at once. The team going backwards already knows what's going to happen, so they can radio tips to the team going > > > forward. Squeezing a problem from both directions — the future helping the past — is exactly what your two-sided method does.
A reusable method for attacking a problem from two directions at once:
- A forward pass works from now: what can I actually do next, with what I have today?
- A backward pass works from done: assume the outcome already happened, write the retrospective, and ask what must have been true — and what nearly killed it.
Neither pass is new on its own (forward planning; backcasting / pre-mortems / Amazon-style working backwards). What this method adds is the loop: the two passes are written independently, then diffed against each other. The gaps found in the diff feed the next round. You repeat until the pincer closes — the forward plan produces everything the backward retrospective requires.
The value lives in the diff. A reasonable forward plan and a plausible backward story will disagree in specific, findable ways — and those disagreements are usually the things that would have sunk the project in week 8, discovered in week 0.
A run produces several files, but you only act on one of them: 40-your-plan.md. It's written action-first — it opens with the 3–5 things to do this week and the single most important insight, then lists the full plan, the risks to watch, and the kill criteria (when to stop or change course). Everything else in the folder (the two passes and their comparison, per round) is the working-out that produced the plan — read it only if you ever want to see why an action is there. Each run folder also gets a START-HERE.md that says exactly this, so you never have to remember which file is which.
So: do the run, then read the plan. You don't need to study the gaps, the wishful-assumption list, or the pre-mortems — the important findings from all of them are already folded into the plan.
You need a pen, the templates in templates/, and 2–4 honest hours. Or run it agent-assisted with the /pincer skill (see below).
-
Frame the problem — copy templates/00-problem-frame.md. Write one sentence: what outcome, by what date, verified how. If you can't write this sentence, stop — the method doesn't fit your problem yet (see "When to use it").
-
Write the forward pass — copy templates/10-forward-pass.md. From today's reality only: next 3–5 concrete moves, resources you actually have, blockers, genuine unknowns. Rule: do not mention the end state. Be deliberately myopic.
-
Write the backward pass — copy templates/20-backward-pass.md. Date it after your horizon and write in past tense — the tense is what shifts your mindset, don't skip it. Cover: what "done" concretely looked like, the last three milestones walking backward ("what had to already be true for this?"), the thing that nearly killed it, and what you'd tell your past self. Rule: do not look at the forward pass while writing. If two people are involved, have them write one pass each — even better.
-
Reconcile — copy templates/30-reconciliation.md. Put the two side by side and extract three lists:
- Missing pieces (gaps) — things the backward pass requires that no forward action produces → these become new forward actions.
- Wishful assumptions (fantasies) — backward claims your forward reality can't support → revise the end state or timeline.
- Already agree (confirmations) — where both agree → leave alone.
-
Loop or stop — if reconciliation found new missing pieces, run another round (a fresh forward pass folding them in, a fresh backward pass with the wishful assumptions corrected). Stop when a round produces no new missing pieces, or only risks you explicitly accept. Typically 2–3 rounds; more than 4 means you're polishing a story, not learning.
-
Write the plan — copy templates/40-converged-plan.md into
40-your-plan.md. It opens with what to do this week and the one key insight, then the full action list, the risks to watch, and the kill criteria (when to stop or change course). This is the only file you'll come back to.
A complete two-round run is in examples/whatsapp-helpline/ — read it first if the steps above feel abstract.
If you use Claude Code, everything above collapses into one command. Open Claude Code in this repo and type:
/pincer launch a WhatsApp helpline for health workers, 500 users in 12 weeks
Claude interviews you once to frame the problem, then runs the whole loop for you: it writes the forward and backward passes using two isolated subagents that never see each other's output (the structural fix for one head writing both passes and agreeing with itself), does the reconciliation in front of you, asks you to confirm or veto the gaps and fantasies, loops, and hands you the converged plan with kill criteria. All artifacts are saved under pincer-runs/<problem-slug>/ so the run is auditable later.
One command, four modes:
| You type | You get |
|---|---|
/pincer <your problem> |
The full facilitated run, 2–3 rounds, files saved |
/pincer quick <your problem> |
One round in chat, ~10 minutes, no files — for smaller decisions |
/pincer resume |
Picks up an unfinished run exactly where it stopped |
/pincer status |
Lists your runs and what each one needs next |
Typical scenarios, verbatim:
/pincer migrate our reporting off the legacy database by end of Q3/pincer get the new grant program running in 3 districts within 6 months/pincer quick should I hire a second developer this quarter/pincer ship v2 of the mobile app with offline mode by December
Use it anywhere: the skill is fully self-contained (it bundles its own copy of the protocol and templates). To make /pincer available in every project, not just this repo, copy one folder:
cp -r .claude/skills/pincer ~/.claude/skills/
Runs are always saved in the project you invoke it from, under pincer-runs/.
Good fit: goal-directed work with a definable outcome but an uncertain path — launches, migrations, grant-funded projects with fixed end dates, org changes, big personal decisions. Especially valuable when external dependencies (approvals, partners, adoption, relationships) have long lead times: those are exactly what the backward pass catches and the forward pass misses.
Poor fit: pure exploration or research where the destination is the unknown (the backward pass degenerates into guessing), crisis response (no time), routine execution (no uncertainty), and open-ended continuous work ("improve culture") unless you first force a concrete horizon onto it.
- Same-head contamination. Solo, both passes come from one imagination with one set of blind spots, and convergence can produce false confidence. Mitigate with the skill's isolated subagents, or a second person.
- Narrative fallacy. The backward pass is a coherent story; coherent is not probable. Treat the output as a risk-finder, not a route guarantee.
- Anchoring. One imagined future can prematurely lock the solution's shape. If the solution space is genuinely open, write two divergent backward passes in round 1.
- Overhead. A round costs real hours. Don't ceremonialize small or well-trodden problems.
PROTOCOL.md the full protocol with rules and stopping criteria
templates/ copy these for each run
examples/whatsapp-helpline/ a complete two-round worked example
.claude/skills/pincer/ the /pincer skill for agent-assisted runs